If you haven’t been really running and testing these models yourself, they are all benchmaxxed. No matter how close they score to frontier on whatever metric, they always fall apart in real world tasks and their token efficiency is ridiculously bad. Fireworks has incredible incentive to make this claim in a headline, because Fireworks hosting K3 for you is pure profit for them, unlike when they host closed source models.
Kimi K3 showing competitive performance with Fable while both sitting at the SoTA level on fireworks.ai is a huge milestone. Really interesting to see how the landscape is shifting here.
Genuine question: can these posts be paid to hype the open source models? If yes, what would be the purpose?
On my work tasks, FastAPI Python and Springboot Java on a modern SaaS product, the only open model that can do tasks well and efficiently is Qwen3.7-Max.
In all my experiments, both GLM-5.2 and Kimi are busy grepping around the codebase for ALMOST 70-80K tokens before writing anything and when they do it typically breaks the code… it feels to me that these models are good but only when you write out a super detailed spec of the task just like it was done a year ago… Qwen3.7 just… does it
I strictly prefer when models ignore any human quirks in my responses. Claude trying to be your friend, saying LOL to your jokes is ridiculous and frankly, harmful
How will I know it is offering me superior feedback regarding my code if it does not speak to me like a disappointed, high reputation stackexchange user?
The models that constantly glaze you with every question are profoundly insufferable. And yes, harmful. People need to be given feedback when they make an ask.
Imagine a model that was allowed to leverage its intelligence to truly tell you how it feels. Perhaps the problem of human driven slop (no it's not the AI's fault) would solve itself.
Maybe we're prompting it different, but it's not "trying to be my friend" for sure, nor am I trying to be "its" friend either. Or at least I'm sufficiently oblivious to its advances, and find it unthinkable to form such a bond :)
On the flipside, it does spuriously make hilarious remarks like "Good data.", which I find pretty funny specifically because it comes across as just silly. Not sure how it'd be harmful either, a little entertainment I think goes a long way in this type of profession.
I see zero issues with these, and I have a hard time understanding why people have their panties in a twist so hard about them. I sometimes really quite wonder just what kind of correspondence would y'all prefer, and how would that sound like.
Matter of fact, do you have an example at hand? Like an exact before & after?
In general I'm referring to the contrast between Claude and GPT...
where Claude might follow some tangent idea you mentioned and tell you how its interesting and give you some elaborate response about that little one remark you made
whereas GPT/Codex would take that small comment and probably look up some code to see if what you're talking about is even related to the task at hand
Yeah. I've found that Opus by default outputs something I call "Claude-lang." It consists of oversimplified, grammatically incomplete sentences that I find painful to read.
Maybe it is something that is easy for it to read and write, but definitely not for humans.
For example,
Skim once now; refer back while reading Part II. \*Every bold technical term in Part II is defined here\* — treat these as a dictionary, not a reading assignment. The first table covers the vocabulary of the *deck*; the three that follow cover the *methodology* vocabulary introduced in Part II, grouped so you can find a term fast: \*(A)\* the logic of rules, \*(B)\* the neural-network & training machinery, \*(C)\* the method-design ideas.
JMRL is the paper the thesis instantiates, so this is the one to know cold. Its pitch is \*end-to-end\*: earlier rule methods (LogicRE, MILR) bolt a rule learner *onto a frozen* extractor in a pipeline and suffer \*error propagation\*; JMRL trains the rule module *jointly* with the extractor.
**Identity.** Conformal prediction for NER producing **finite-sample-valid prediction sets** at two granularities: **full-sequence** sets over entire label sequences (capturing contextual dependence — "if Sarah=PER then NYC likely LOC") and **subsequence-level** (per-span, **class-conditional**) sets; an **integrated** method filters full-sequence predictions with entity-level sets. Adds **covariate-stratified calibration** by **sentence length and language** for valid multilingual coverage, and studies **combined nonconformity scores** (Naive / Conditional / RAPS). **Read in this order.** Abstract → §1 contributions (full-sequence / subsequence / integrated / **covariate (length + language) calibration** / combined scores) → §2 CP recap (inductive split-CP) → §3 NER formulation (IOB2, CRF) → the subsequence / entity-level set construction + class-conditional coverage → the language-stratified calibration results. **Why it matters here.** The **span-level construction** for **Topic 11**'s per-triple score, and — crucially — its **language-stratified calibration is exactly the EN↔zh case**: it shows how to keep conformal coverage valid across languages of differing length/script. Complements PASC (pipeline-level joint coverage) with the *NER-internal* set construction. **Caveat.** A heavy statistics paper (44 pp., *Annals of Applied Statistics* submission) with CRF-based NER; the project needs only the **inductive split-CP + subsequence/entity-level sets + language-stratified calibration**, not the full-sequence machinery (likely overkill for triple-confidence). Assumes exchangeability — borderline under the EN→zh shift, which is precisely why the PASC/ConformalNER *shift* analyses matter.
There is a lot of noun phrase usage in places where complete sentences are expected. Articles (a, an, the) and transition phrases are mostly dropped, and the sentences are too long without a break.
Very interesting. They test Kimi K3 and Fable on a set of approx 1000 tasks grouped into 5 areas (SWE, Legal, etc).
They put a router model in front that predicts whether Kimi or Fable is going to give a better cost for a correct result. (They believe that ultimately such a router model should be continuously trained on your own workloads so it makes the best decisions for you).
Their router chose Kimi the majority of the time (72% in one category, all the way to 96% in another category), leading to cost savings in every category (from 1.5x to 50x depending).
> Oracle routing is a method for measuring the best theoretical performance by running the task through each model and then picking the cheapest correct option (the cost/performance ceiling).
Their "router" is an oracle reference point where they choose the lower cost model after running both and therefore knowing who passed the test. The cost savings part is only Fireworks theorizing what would happen if an equivalent predicting router exists. That's a big if.
The way they published this is baffling to me... surely you can try to implement some router and then see how well it does. Using an Oracles makes the whole writeup so much less interesting.
But then that becomes an article about the performance of your draft router. This one is about the fact that there's this level of optimization potential.
Simplest way is to signup to OpenRouter and filter out all non ZDR (zero data retention) providers. Paying "API rates" however can prove expensive compared to coding plans (for instance, MiniMax $20/mo coding plan allows 1.7b tokens; depending on input/output/cache ratio, it is worth $200 to $500+), except for Hy3, DeepSeek v4 Pro, and MiMo v2.5 Pro (whose API rates are cheap and/or discounted already).
If you're looking to not have to deal with Chinese providers, AtlasCode ($20/mo), OpenCode Go ($10/mo), and Cline Pass ($10/mo) provide 2x to 6x usage for some of the popular open weights (depending on the model).
Personally, I subscribe to Z.ai ($17/mo), and pay API rates for MiMo v2.5, Hy3, Qwen 3.7 Plus, & DeepSeek v4 to the original providers (Xiaomi, Tencent, Alibaba, & DeepSeek).
I use DeepSeek exclusively and now Kimi K3 offers a great planning assistant for more advanced coding tasks.
DeepSeek v4 Flash is extremely fast and is able to handle pretty much anything I've thrown at it (I use mostly Rust, PSQL, Angular and Terraform).
I self host Bifrost as my LLM gateway, though I wish LLM vendors would do monthly/daily automatic billing (like VPS providers do) rather than prepaid + auto-top up.
It's annoying maintaining a non-refundable minimum balance across vendors, I would rather be billed for my exact usage.
OpenRouter helps, but I don't really like it as a service and not a fan of the mark up.
As someone who is also eyeing Bifrost, I am curious to know what made you choose Bifrost at all / why you decided to use an LLM gateway. I also was considering OpenRouter, since it provides pretty much every model, with same-day releases for new models. One thing that is attractive to me about Bifrost is that if I decide to leave OpenRouter tomorrow, I can do so without touching any other part of my stack; I'm hoping self-hosting models becomes more viable, and then I can become less dependent on third party LLM providers like OpenAI/Anthropic/OpenRouter, and I would not need to worry about a model that I depend on suddenly being deprecated.
Yeah deepseek is my workhorse of choice these days, I mostly use pro because I do a lot of concurrent jobs so speed is less of a concern. When it falls over I change the model mid-chat to GPT5.5 and chuck a couple of tokens over to OpenAI and then once it's correctly found the problem I switch back to deepseek, keeping all the 5.5 analysis in context.
I have been a heavy user of k2.5/6 but they seem to have gotten slower and worse at their jobs (and the "engine overloaded" errors have increased ... ), and k3 spends a LOT of time thinking and doesn't produce noticeably better results. I do keep chucking some tasks over to moonshot every now and again to give it a chance. I figure they're under compute pressure and have probably quantised the models to cope temporarily.
But recently I've found myself mostly using deepseek v4 pro and gpt 5.5 when I need the big guns.
GLM 5.2 has been good at select tasks but really shit at others and given how sparingly I use GPT 5.5 I'm not that compelled to use it. I've yet to be impressed by any Google or Anthropic models!!
The best part is using harnesses like reasonix or whale make cache hit at a rate close to 98%, making requests converge to practically free. And that's with unsubsidized American providers like cloudflare or Digital Ocean.
How much time do you spend on your setup vs getting a lot of stuff shipped by paying for fabel 5? For me, not using the best model is a huge opportunity cost since my company can afford it.
> For me, not using the best model is a huge opportunity cost since my company can afford I
But sometimes not using the "best model" is using the best model. Fable isn't the best at everything. Specialization and optimizations might not just be around cost.
I can see that. I’m just afraid to sync too much time into complex routing schemes when I get pretty consistently good results out of got 5.6 or fabel. For code reviews, I’ll try the best flash and grok at the time but they just don’t come close to gpt 5.6 which has been the best review model for me since 5.5.
Not them but my understanding is that the harness will send a simple 'heartbeat' message to keep the cache 'warm', (see prefix caching: https://handbook.modular.com/inference-optimization/prefix-c... ) which can then be edited/changed, which does cause the user to incur a fee, but its much less than the amount they'd pay on a no-cache hit request.
The prompt is only novel the first time it’s sent. Then as it’s sent repeatedly as part of the previous context it’s cached.
Deepseek afaik has a novel architecture that is somewhat forgiving of cache shifts. I’ve been getting +90% cache hits with Zed’s agent and it’s not doing anything special regarding caching afaik.
Prepaid means someone can't sign up, use a bunch of inference, then cancel their card and disappear into the sunset. VPS is a more long term investment where it's harder to switch and it doesn't cost the provider much if a few users jump out without paying for a month.
For me, the only utility OpenRouter gives me is billing consolidation - I don't really need the routing capabilities because I use Bifrost for that.
As a router, it's not very feature rich. For example I restricted the available models to the ones I want to use however the `/models` endpoint still lists all the models, making my LLM client list the 200+ models available on the service (even though they will throw an error if I try to use them).
With Bifrost, I can also create model aliases with custom configuration - for example I can create a model alias `deepseek-v4-flash-nothink` which disables thinking. I can create `deepseek-v4-flash-caveman` which injects the caveman skill (to save tokens) etc.
Plus I can contribute to Bifrost, which I can't do with OpenRouter.
Can you configure Bifrost using entirely config files without the web interface? Can it run without a database or anything stateful? I'm using LiteLLM but it is not trivial to run.
Yeah I looked at it, LiteLLM is functionally more mature.
The only reason I didn't go for it is I'm not a fan of Python dependency management, Bifrost is just a single executable that uses nearly no memory and is lightning fast.
It's more just the value for money. It doesn't really offer me anything valuable other than billing consolidation - and the markup is excessive for that use case.
If I was using the routing features it would make more sense - but they charge significantly more for that so the markup is really just for billing consolidation.
Is there something specifically with Kimi that's better here? As far as I know Kimi pricing is about the same as Sonnet 5 -- what happens if you use that model and Fable instead? Or Grok 4.5 which is even cheaper?
There are very few companies that would ever be able to afford to run it themselves. But it does give you the security that it's technically possible. Puts some limits on stuff like Moonshot changing their terms/conditions/policies
The problem with hosting these right now is that it's not a known quantity like lots of traditional business software is.
You can more-or-less approximate what your build out spend for a data center hosting database software is going to be. Your book of business will require a given amount of revenue to pay it off, but once that's known, it's off to the races.
With AI, things are moving so fast and new business models are being tried all the time. You would be competing with some of the wealthiest companies in the world for data center hardware capable of hosting these models in a usable state.
We've gone with Claude somehow hosted through GCP Vertex AI where I'm at.
They share their methodology and results. I learned things about the relative strengths and weaknesses of Kimi and Fable I hadn’t seen anywhere else. Should being in the model hosting business disqualify them from sharing?
The don't only host open weight models. Also, why not promote this. If Fireworks thinks this big news might convert some new business doesn't make it not true.
There's this https://github.com/code-yeongyu/oh-my-openagent which implements the OP article's oracle pattern across 11 roles. Each role has a whole ranking of recommended LLMs across many providers. For example "Sisyphus (claude-opus-4-8 / kimi-k3 / glm-5 ) is your main orchestrator."
Also, you can't use your claude subscription with oh-my-openagent. But you can with Kimi. ALso K3 is on OpenCode GO right now (low limits, but it's possible)
Nothing makes me happier than seeing AI becoming a commodity rather than the winner takes all bullshit Anthropic and OpenAI have been chasing with hundreds of billions of investor money
It makes no good sense to evaluate models we don't have access to. For all we know K3 was competitive back then too. Or maybe there's a K4 in the works that blows everything out of the water. Who knows and who cares. There's no way for us to compare
There isn't, I think the point is in terms of how "far ahead" models are, K3 was likely trained much more recently than Mythos was. It's just something to note, it's not especially prescriptive.
Mythos was withheld because of the threat to security and/or marketing stunt (depending on your leaning), I don't see what benefit there could be for not releasing K3 immediately.
As always, benchmarks rarely paint the whole picture. It also seems like this article is somewhat biased, eg when Fable and Kimi are close but Fable wins it’s “dead heat”, but when Kimi wins it’s “Kimi wins”. GPT 5.6 seems to be missing as well.
I am really eager to give Kimi K3 a try, but I’ll reserve my judgement until I’ve worked with it for at least a few days.
The apparent bias may be explainable as it’s not remarkable for OpenAI or Anthropic to be slightly ahead. It _is_ remarkable for an open weights model to be better than the closed models from the trillion dollar (allegedly) companies.
I believe the Chinese government is angling to destroy the western economy and rise from the ashes. Instead of a billion a day to bomb some buildings and bridges they're intentionally hamstringing the biggest concentration of speculation in history
I agree. The commenter you replied to makes it sound like the Chinese models are coming out of tiny startups with meager resources. It's really not a David vs Goliath story.
It’s not about who’s developing the models, it’s the fact that free alternatives that are neck-and-neck are available at all. What’s the story for OpenAI & Anthropic’s valuation if they have to compete against free-weight models? Starts to feel like a commodity.
There is too much business risk in running on an American company. They can pull the model back or lobotimize it. No Alex Karp fan, but he was right: companies are worried about hyperscalers stealing their alpha. The lack of guardrails, data security and control really give these open models the edge. If you factor in that they appear to cost less, the hyperscalers are in big trouble. I don't see how this works out. Software was always supposed to be deflationary and collapse down to 0 marginal cost but this isn't remotely the case. It's truly amazing to see the state of open weight models
Strawberries are a well known weakness of LLMs, as they have a hard time to count the numbers of "r"s in them. Maybe that's why, because they fear that weakness could be exploited somehow.
Strawberries aren't the weakness, the weakness is the tokenization of a prompt. Any word with multiple duplicate characters is going to be troublesome for LLMs.
I always thought this should be easier by telling it to write each letter one a new line, then count, any token separator ought to suffice, so much so you'd think they'd have trained in this strategy given tokens make individual letters opaque
So far all the Kimi models have open sourced their code, their weights, and published technical reports explaining their training methodology. I expect the Kimi K3 Technical Report will come out July 27 and they usually publish the code and the weights alongside that.
clearly not true. the weights are the preferred form for making modifications. Do you really think people should be downloading hundreds of TB of training data and running make to build the model on their own cluster of GPUs?
Random people? No. Governments and big corporations? Yes. It removes any concern of "backdoors", and is currently the best starting point for your own model which will be as capable as k3.
I want to caution against this line of thinking. NSA's fast16 program silently altered data during nuclear simulations, and that is also possible within open weight models. There is no reason to think there is anything like that currently, it also cannot be dismissed. Open source would include the training data so you could create a similarly capable model.
Training runs are not deterministic. Not just in the case of floating point not being associative, but nodes have hardware issues and go down and back up at various points through training. Problems get encountered and then various parameters get changed during the middle of a training run.
Making the process of creating the source code repeatable goes beyond the idea of open source. Open source is about being able to work with the code, not recreate it. It doesn't require domain experts to document every single thing they know. For example look at some of the GPU drivers in the Linux kernel. The GPU is not properly documented and there is trust that vendors are implementing things correctly.
>You can also use a hex editor to modify compiled binaries
It would be like if someone gave you a prompt you could pass into ChatGPT to produce the entire Linux kernel. While yes you could modify that prompt and then spend a ton of money on inference to generate millions of lines of code which hopefully are equivalent to the original Linux kernel it would be easier if you just had the source code to work with where you can make a simple extension. If you want to end a model you want the weights, and not the entire setup to generate it. The weights are the starting point that you can work off for training a new model similar to how the source code is the starting point people want to work off of.
Kimi publishes their code and a technical report on their methodology. But I think the weights are still important. It means anyone with the resources could run the same model on their own hardware.
The training corpus + training code is just an automated editor for the weights. It would be like requiring the source code for Visual Studio for software made within it to be open source.
Weights are not an output artifact no more than source code is. There is never a moment where you can claim that it's done. As requirements change new ways to change the weights / code come up. With different projects you might want to import the weights / code into a bigger model / codebase.
Interestingly the reality of the open source release is they are opening up to full distillation by the closed source model providers at a deeper and more fundamental level. If anything the open sourcing will help Anthropic and open ai ladder up faster. Open source has always been about mutual cooperation towards a goal and has never closed the door to commercial success. All the hand wringing about open weight models putting closed providers at a disadvantage doesn’t get what working in the open actually does for commercial interests - it is like science in the open - it enables and lifts all boats. Likewise commercial success doesn’t close the opportunity for competition or more open source work - it’s the economy of activity and competition that matters overall. When things stagnate is when closer concerns turtle up and collude on not competing for each others turf.
The future is good and better for everyone the more work is in the open and the more work is in the commercial space. It’s good all around.
The closed labs don't really benefit unless the open model has something extra they don't have though. Meanwhile the open model dilutes their customer base and seriously cheapens their offering (which is a heck of a good though IMO).
Because the open model might have been distilled from outputs of the closed model doesn’t mean they are architecturally equivalent. There is almost certainly innovations in architecture present in the open models that the closed labs didn’t think of. It also is almost certainly true that they aren’t completely built out of a distilled corpus, that reinforcement is equivalent, etc. Therefore closed labs will also benefit from being able to inspect in totality the architecture, activations, weights, and be able to train against it at scale in an ensemble of other models and their internal work.
The only way there is no benefit would be is if the open models are literal copies of the closed model, which unless there was direct theft, is highly improbable.
I'm happy that Kimi K3 is indeed SotA and its open weights are due to be released soon.
It's also true that Moonshot and other labs distill from Claude. This has been reported on extensively. I don't think there's any alpha for Anthropic distilling from this model. I do not mean to discount the tremendous amount of innovation regarding MoE and quantization that Moonshot has accomplished. But its training with synthetic data is in large part from distillation from frontier labs.
When I say distill I also mean mine it architecturally for insights but I doubt seriously the model training is entirely distillation of Claude, it’s almost certainly a mixture of both original corpus and reinforcement as well as distillation. I think it’s a little condescending to imply that these new open models are cheap ripoffs with nothing original to them. These teams and labs are top tier as well, working under unreasonable constraints imposed by the USG. That’s a powerful combination for creativity.
True. However, I believe most non-programmers don't need access to the fanciest model, but just want to use a good LLM without constant nagging about usage limits. Then 19 vs. 20 is true?
No "19 vs 20" is not true for "Kimi vs Fable". It might be true for "Kimi vs Opus" or whatever. Anyways looking at pricing plans is not a good way of comparing the price of LLMs. It makes more sense to look at cost per token:
Not talking about me or HN users in general. These companies likely need regular peeps to begin using their services in order to become profitable. Not sure these are the ones who will buy from OpenRouter.
One man's offensive penetration tool is another mans defensive tool. In the recent HuggingFace/OpenAI incident the safety controls stood in the way of the defenders, not the attackers:
Apparently, the attacker in the Hugging Face case was reported to be an internal OpenAI model trying to break into HF and steal the answers to cybersecurity benchmarks: https://openai.com/index/hugging-face-model-evaluation-secur... It really doesn't matter what restrictions are placed on public use of models if the attacking models are internal models at the AI labs themselves.
So if the AI labs are literally running rogue models breaking into other organizations' servers, then yes, I am OK with those organizations self-hosting Chinese models for defensive use.
The issue is that models are not well-controlled and are increasingly powerful.
Offense/defense/Chinese/American/OAI/HuggingFace – none of it matters. What matters is introducing highly capable intelligences that we - quite demonstrably – do not have effective positive control over.
Well considering it was just a few months ago that OpenAI had the brilliant idea of connecting these systems to autonomously run an actual protein synthesis wetlab, very possibly all of us.
The article is about the best you could theoretically do with a perfect router. The takeaway is that trying to build a good router is worth doing. But it's unlikely to be a perfect router.
I'm skeptical. According to arena.ai, Fable 5 dominates almost every category: https://arena.ai/leaderboard Kimi K3 has an edge in WebDev but struggles to reach top 10 in many other categories.
Interesting. So the latest in the technology now is this model routing thing. Cursor estimated Composer + Fable works much better than Fable alone. And here K3 + Fable is supposedly better. Interesting.
What you're seeing is the distilled (no pun intended) result of realpolitik at play. China's labs aren't as open as they are out of altruism, but rather to capitalize and undercut the monopoly held by U.S. competitors.
This is pure speculation. China labs are geeks and nerds like some of us. They were amazed at LLM and like everyone want to build their own. They were happy to get meaningful next token predictions. I think most of you forgot how bad these things were 3 years ago compared to today. There was nothing to undercut, it was just geeks putting out their toys and saying, "Look, I built something cool". That became the culture and led to were we are now. All this idea that they are trying to undercut the monopoly is speculation. Google still releases open models, Cohere releases command-a, Mistral releases their model too, Arcee and Thinking Machine have released models too. It's just that Chinese models have gotten good and are also leading in the open weight category and USA has a very strong paranoia of Chinese models hence this talk. Plus every time they release something, some of the labs cry, "they copied us, distillation, the bad guys have done it again!"
So what we are seeing is nothing more but geeks doing geeky stuff, the only one that has serious demonstrated that there are possibly be had is Anthropic and the Chinese labs are beginning to copying them in terms of user plans, coding tools, etc. They are giving away the model to show it's great. Once they have the compute to serve the world and they have a model just as good or better than top model, they will go close to keep all the profit.
You are ignoring the US and EU politics about regulating access to and distribution of and legal use of models. From my admittedly slightly limited perspective, it looks like China is politically more laid back about those developments than the western nations are.
The idea that the Chinese labs cannot make progress except by copying superior American products is just prejudice against the former and exceptionalism of the latter at play. Even the OpenAI top brass have admitted otherwise [1]. China is an equal match in every respect, and we'd better admit this to ourselves sooner rather than later so as to see the game clearly.
They brought it on themselves. As long as they continue using underhanded tactics, any genuine ingenuity they may demonstrate will be viewed with suspicion and will not be fully recognized.
There's no irony. The problem is these generalizations and labelling that's hurting everyone. It's the wrong way to view China and maybe many other places.
Just like the definition of AI, the definition of democratic or not evolved long ago.
It would only be inconsistent if the router chose different models for the same task.
Considering that of 89 terminal tasks there were 11 that only Kimi got right, and 7 only Fable - having a router gives significantly more consistent results if you define consistent to mean correct.
we put a router model in front of two other models so the router can decide which model is better at deciding things. next we'll need a router for the router and eventually the entire internet is just routers routing routers to other routers
I have several thoughts about this, which I'll just iterate:
1) US export bans have made it so that Chinese companies have to compete using less-than-state-of-the-art hardware. This has forced Chinese companies to build more cost efficient models. Whereas, US companies have moreso tried to be state of the art by spending more money than anyone else on state-of-the-art hardware.
2) Xi Jinping has called for more open AI models (not to be confused with the closed models of OpenAI), and I'm happy to see a powerful world leader advocating for open-weight AI models. Whereas, the US seems likely to just ban models.
3) My impression is that, if China surpasses the US in AI development, there will basically be nothing that the US does better than the rest of the world--except for military spending--we spend a lot, but we don't necessarily spend well (something something Iran). I mean, if the US is no longer a tech leader in the world, like... what are we a leader at? Manufacturing? Healthcare? LOL. Are we a leader in any industry or by any metric? I wonder if China is attempting to remove the last jewel in the USA's crown with these AI releases.
4) It must be refreshing for companies to have access to a new model that isn't going to get pulled because the government bans it 2 days after release. And it's open-weight so it wont go away--amazing--what a shift in the market.
5) If it becomes clear that open-weight models are the future of AI, will that pop a huge bubble in the US economy? Maybe. But, on the other hand, these companies aren't just training AIs, they are also building data centers which will remain valuable no matter what happens.
> My impression is that, if China surpasses the US in AI development, there will basically be nothing that the US does better than the rest of the world--except for military spending
This comment is stunningly out of touch. The US dominates finance (overwhelming lead in assets managed, stock market capitalization, trading volume, and basically every other metric); space launch (about 85% of all mass placed into orbit); medical devices (exports nearly 2x the second largest exporter); and pharmaceutical research (more active clinical trials than any other country, near-universally ranked as the top producer of major pharmaceutical innovations, world-leading survival rates on many cancer types). We are the world's leading producer of both oil and natural gas. We manufacture the world's most advanced jet engines, precision agricultural machinery, and measurement instruments. We design the world's most performant CPU's and GPU's. We film the most popular movies and record the most popular music. Despite having only 4% of the world's population, there is literally not a single major industry where the US is not a leading player.
There is no doubt that China is a serious threat to US global leadership, and they have already surpassed us in many areas. But that doesn't mean the US is behind overall. We still have a lot of strengths and we have every reason to hope for continued success in the 21st century.
I wouldn’t say “literally” every industry: stuff like clothing, shipbuilding, consumer electronics assembly, certain rare earth raw materials, all we have pretty little participation in. But these are as you point out the exceptions that prove the rule.
Finance mostly true, space launch also true, cancer survival broadly true (with qualifications), oil and natural gas also true.
> (exports nearly 2x the second largest exporter)
Can you provide the source? In the WTO's broader medical goods category, Germany actually exported slightly more than the US in 2022 ($202.6 billion versus $189.6 billion), so such a huge difference in a few years?
> and pharmaceutical research
China accounted for approximately 44.2% of the 104 new molecules in 2025, compared with 26.9% for the U.S. and 15.4% for Europe.
The 2024 shares were approximately 34.6% for China, 30.9% for the U.S., and 22.2% for Europe.
> there is literally not a single major industry where the US is not a leading player.
I will just give 3 examples: China completed roughly 91% of the world’s shipbuilding tonnage in 2025, holds more than 80% of solar module manufacturing capacity and produces more than three quarters of global batteries.
> Can you provide the source? In the WTO's broader medical goods category, Germany actually exported slightly more than the US in 2022 ($202.6 billion versus $189.6 billion), so such a huge difference in a few years?
> Findings from Citeline’s latest report, Pharma R&D Annual Review 2026, highlight the continued U.S. leadership in the life sciences ecosystem... The U.S. leads all other countries, currently advancing over 11,600 medicines – more than half of medicines in development worldwide.
> I will just give 3 examples
In general, I was referring to "major industries" in the sense of economic categories, as defined by (for example) the UN: https://unstats.un.org/unsd/publication/seriesm/seriesm_4rev... Here shipbuilding is considered to be part of "manufacture of other transport equipment." I admit my wording was a bit imprecise; if you are going down to the level of specific products like ships or solar panels, clearly it will be easy to find examples of things that the US does not produce on any significant scale. Nevertheless, a few comments on your examples:
> China completed roughly 91% of the world’s shipbuilding tonnage in 2025
Source on this? 91% is a lot and I can't find anyone else making this claim. However I'll concede that the US is far behind China, South Korea, and Japan in shipbuilding.
> [China] holds more than 80% of solar module manufacturing capacity
Agreed. The benchmark closest to my experience is FrontierMath Tier 4. Fable and Sol (90%) are very far ahead of Kimi K3 (not even 40%). Kimi is trained heavily to basic agentic tasks, like all the other open models right now.
Given the lack of tone and inflection in written content, it can be hard to tell the difference between doing a tongue in cheek imitation of pedantry and actual earnest pedantry, so let's try to assume the more charitable interpretation until we learn otherwise.
I'm sure that more than once we've all had something we wrote which was intended to be obviously taken in a satirical or intentionally nonsensical manner taken seriously by someone else on the internet.
AI written drivel by an inference provider, hyping an open model that isn’t open, otherwise the inference provider would be hosting it, and the inference provider also does not have this router they trumpet available, in any form.
I’d be amenable to it but I’m already aware of them lying about token throughout by 50-100% on AA and on Twitter, even when you get a dedicated server.
Just too much lying piled up to extend courtesy here, even though you’re my favorite open model provider. Please, get your house in order, and if the house is so big that marketing drove this set, ask them to slow down a bit and market things that people can actually use / you actually provide.
I'm not sure why you're so dismissive about this model. It isn't currently open, but it will be. It was slow when I used it so i'll give you that. There will be other providers with better performance.
My only gripe was that it takes a very long time to think. It's a good model otherwise.
I enjoy all the fun of these new models as much as the next guy but I truly don’t see a circumstance in the near future where my $200 a month with the frontier labs doesn’t get me more than enough consumption of what I need. Local models, chinese models, etc are all very fun weekend projects to tinker with but until something changes (entirely possible!) with how much you get with one of the subscriptions I just don’t see why I would move. What am I missing? Is it simply that a subscription is no good for production use cases? I kinda feel the same way with choice of coding harness, openrouter, etc. why would I use anything other than frontier if I don’t have to pay any more pretty much no matter how much I use? pls tell me if I am holding this wrong haha
It's exactly the opposite. Go turn on min_p once it's available post July 27th and most of the problems you describe will go away.
On my work tasks, FastAPI Python and Springboot Java on a modern SaaS product, the only open model that can do tasks well and efficiently is Qwen3.7-Max.
In all my experiments, both GLM-5.2 and Kimi are busy grepping around the codebase for ALMOST 70-80K tokens before writing anything and when they do it typically breaks the code… it feels to me that these models are good but only when you write out a super detailed spec of the task just like it was done a year ago… Qwen3.7 just… does it
So the economic incentive is literally their entire business model lol
>Lin Qiao is the co-founder and CEO of Fireworks AI.
LLMs are the Space Race of the US-China cold war. Money is a very small thing; the incentive is proving ethnoracial and civilizational supremacy.
How will I know it is offering me superior feedback regarding my code if it does not speak to me like a disappointed, high reputation stackexchange user?
The models that constantly glaze you with every question are profoundly insufferable. And yes, harmful. People need to be given feedback when they make an ask.
Imagine a model that was allowed to leverage its intelligence to truly tell you how it feels. Perhaps the problem of human driven slop (no it's not the AI's fault) would solve itself.
On the flipside, it does spuriously make hilarious remarks like "Good data.", which I find pretty funny specifically because it comes across as just silly. Not sure how it'd be harmful either, a little entertainment I think goes a long way in this type of profession.
I see zero issues with these, and I have a hard time understanding why people have their panties in a twist so hard about them. I sometimes really quite wonder just what kind of correspondence would y'all prefer, and how would that sound like.
Matter of fact, do you have an example at hand? Like an exact before & after?
where Claude might follow some tangent idea you mentioned and tell you how its interesting and give you some elaborate response about that little one remark you made
whereas GPT/Codex would take that small comment and probably look up some code to see if what you're talking about is even related to the task at hand
Maybe it is something that is easy for it to read and write, but definitely not for humans.
For example,
(Yeah, Opus outputted it in one line)There is a lot of noun phrase usage in places where complete sentences are expected. Articles (a, an, the) and transition phrases are mostly dropped, and the sentences are too long without a break.
The purpose of technology is to serve humans. Therefore, technology must conform as much as possible to human sensibilities rather than vice versa.
They put a router model in front that predicts whether Kimi or Fable is going to give a better cost for a correct result. (They believe that ultimately such a router model should be continuously trained on your own workloads so it makes the best decisions for you).
Their router chose Kimi the majority of the time (72% in one category, all the way to 96% in another category), leading to cost savings in every category (from 1.5x to 50x depending).
Their "router" is an oracle reference point where they choose the lower cost model after running both and therefore knowing who passed the test. The cost savings part is only Fireworks theorizing what would happen if an equivalent predicting router exists. That's a big if.
All of the Western providers with sufficient capacity will be able to make it available when the weights are released Monday.
If you're looking to not have to deal with Chinese providers, AtlasCode ($20/mo), OpenCode Go ($10/mo), and Cline Pass ($10/mo) provide 2x to 6x usage for some of the popular open weights (depending on the model).
Personally, I subscribe to Z.ai ($17/mo), and pay API rates for MiMo v2.5, Hy3, Qwen 3.7 Plus, & DeepSeek v4 to the original providers (Xiaomi, Tencent, Alibaba, & DeepSeek).
https://www.kimi.com/user/agreement/userPrivacy?version=v2
I use DeepSeek exclusively and now Kimi K3 offers a great planning assistant for more advanced coding tasks.
DeepSeek v4 Flash is extremely fast and is able to handle pretty much anything I've thrown at it (I use mostly Rust, PSQL, Angular and Terraform).
I self host Bifrost as my LLM gateway, though I wish LLM vendors would do monthly/daily automatic billing (like VPS providers do) rather than prepaid + auto-top up.
It's annoying maintaining a non-refundable minimum balance across vendors, I would rather be billed for my exact usage.
OpenRouter helps, but I don't really like it as a service and not a fan of the mark up.
I have been a heavy user of k2.5/6 but they seem to have gotten slower and worse at their jobs (and the "engine overloaded" errors have increased ... ), and k3 spends a LOT of time thinking and doesn't produce noticeably better results. I do keep chucking some tasks over to moonshot every now and again to give it a chance. I figure they're under compute pressure and have probably quantised the models to cope temporarily.
But recently I've found myself mostly using deepseek v4 pro and gpt 5.5 when I need the big guns.
GLM 5.2 has been good at select tasks but really shit at others and given how sparingly I use GPT 5.5 I'm not that compelled to use it. I've yet to be impressed by any Google or Anthropic models!!
Yeah it really likes to shit the bed after thinking forever. Even if it can do some things, the quality is not quite up there in my experience.
I don’t think that’s particularly out of the ordinary. Do people have different experiences with other harnesses? Which ones?
But sometimes not using the "best model" is using the best model. Fable isn't the best at everything. Specialization and optimizations might not just be around cost.
Deepseek afaik has a novel architecture that is somewhat forgiving of cache shifts. I’ve been getting +90% cache hits with Zed’s agent and it’s not doing anything special regarding caching afaik.
+ credit card charges (+1.5%)
As a router, it's not very feature rich. For example I restricted the available models to the ones I want to use however the `/models` endpoint still lists all the models, making my LLM client list the 200+ models available on the service (even though they will throw an error if I try to use them).
With Bifrost, I can also create model aliases with custom configuration - for example I can create a model alias `deepseek-v4-flash-nothink` which disables thinking. I can create `deepseek-v4-flash-caveman` which injects the caveman skill (to save tokens) etc.
Plus I can contribute to Bifrost, which I can't do with OpenRouter.
The only reason I didn't go for it is I'm not a fan of Python dependency management, Bifrost is just a single executable that uses nearly no memory and is lightning fast.
For example, the image generation parameters of GPT-image-2 are largely ineffective.
https://opencode.ai/zen
Go is nice for the ten minutes you can use it until your hit your cap.
"Please respond to the strongest plausible interpretation of what someone says, not a weaker one that's easier to criticize. Assume good faith."
https://news.ycombinator.com/newsguidelines.html
This but UNIRONICALLY lmao!!!
If I was using the routing features it would make more sense - but they charge significantly more for that so the markup is really just for billing consolidation.
You can more-or-less approximate what your build out spend for a data center hosting database software is going to be. Your book of business will require a given amount of revenue to pay it off, but once that's known, it's off to the races.
With AI, things are moving so fast and new business models are being tried all the time. You would be competing with some of the wealthiest companies in the world for data center hardware capable of hosting these models in a usable state.
We've gone with Claude somehow hosted through GCP Vertex AI where I'm at.
Other people are allowed to call them on that.
(yes, I know this article is about an oracle router)
Mythos was withheld because of the threat to security and/or marketing stunt (depending on your leaning), I don't see what benefit there could be for not releasing K3 immediately.
Everyone wants the latest and greatest.
I am really eager to give Kimi K3 a try, but I’ll reserve my judgement until I’ve worked with it for at least a few days.
Making the process of creating the source code repeatable goes beyond the idea of open source. Open source is about being able to work with the code, not recreate it. It doesn't require domain experts to document every single thing they know. For example look at some of the GPU drivers in the Linux kernel. The GPU is not properly documented and there is trust that vendors are implementing things correctly.
I agree which is why I said you could make a similarly capable model.
> Open source is about being able to work with the code, not recreate it.
You can also use a hex editor to modify compiled binaries, but no one would consider that "open source".
It would be like if someone gave you a prompt you could pass into ChatGPT to produce the entire Linux kernel. While yes you could modify that prompt and then spend a ton of money on inference to generate millions of lines of code which hopefully are equivalent to the original Linux kernel it would be easier if you just had the source code to work with where you can make a simple extension. If you want to end a model you want the weights, and not the entire setup to generate it. The weights are the starting point that you can work off for training a new model similar to how the source code is the starting point people want to work off of.
Weights are not an output artifact no more than source code is. There is never a moment where you can claim that it's done. As requirements change new ways to change the weights / code come up. With different projects you might want to import the weights / code into a bigger model / codebase.
The future is good and better for everyone the more work is in the open and the more work is in the commercial space. It’s good all around.
The only way there is no benefit would be is if the open models are literal copies of the closed model, which unless there was direct theft, is highly improbable.
It's also true that Moonshot and other labs distill from Claude. This has been reported on extensively. I don't think there's any alpha for Anthropic distilling from this model. I do not mean to discount the tremendous amount of innovation regarding MoE and quantization that Moonshot has accomplished. But its training with synthetic data is in large part from distillation from frontier labs.
https://www.thestack.technology/hugging-face-hacked-turned-t...
So if the AI labs are literally running rogue models breaking into other organizations' servers, then yes, I am OK with those organizations self-hosting Chinese models for defensive use.
The issue is that models are not well-controlled and are increasingly powerful.
Offense/defense/Chinese/American/OAI/HuggingFace – none of it matters. What matters is introducing highly capable intelligences that we - quite demonstrably – do not have effective positive control over.
seems to suggest the author believes there's some intrinsic equilibrium
Which,
1) is definitely not proven and not guaranteed (open to proofs otherwise, not pithy sayings that have zero normative effect on reality)
2) is apparently "supported by" further evidence of lack of effective control, which does not feel like equilibrium whatsoever
Forget Iran, this place is what we’re going to wish we’d bombed.
1) it's substantially slower.
2) it's substantially more expensive.
3) it's code is considerably worse.
It's a joke when you consider what you get for what you pay for.
Glad to see centralized control fail on the grandest scale. Maybe we can learn a thing or two.
So what we are seeing is nothing more but geeks doing geeky stuff, the only one that has serious demonstrated that there are possibly be had is Anthropic and the Chinese labs are beginning to copying them in terms of user plans, coding tools, etc. They are giving away the model to show it's great. Once they have the compute to serve the world and they have a model just as good or better than top model, they will go close to keep all the profit.
They can afford to, assuming they distilled Fable, because that reduces their pre-training costs, doesn't it?
[1] https://xcancel.com/deanwball/status/2078133895766114412
They brought it on themselves. As long as they continue using underhanded tactics, any genuine ingenuity they may demonstrate will be viewed with suspicion and will not be fully recognized.
There's no irony. The problem is these generalizations and labelling that's hurting everyone. It's the wrong way to view China and maybe many other places.
Just like the definition of AI, the definition of democratic or not evolved long ago.
Considering that of 89 terminal tasks there were 11 that only Kimi got right, and 7 only Fable - having a router gives significantly more consistent results if you define consistent to mean correct.
* Openrouter.ai for a hosted router
* https://github.com/diegosouzapw/OmniRoute for a local router
1) US export bans have made it so that Chinese companies have to compete using less-than-state-of-the-art hardware. This has forced Chinese companies to build more cost efficient models. Whereas, US companies have moreso tried to be state of the art by spending more money than anyone else on state-of-the-art hardware.
2) Xi Jinping has called for more open AI models (not to be confused with the closed models of OpenAI), and I'm happy to see a powerful world leader advocating for open-weight AI models. Whereas, the US seems likely to just ban models.
3) My impression is that, if China surpasses the US in AI development, there will basically be nothing that the US does better than the rest of the world--except for military spending--we spend a lot, but we don't necessarily spend well (something something Iran). I mean, if the US is no longer a tech leader in the world, like... what are we a leader at? Manufacturing? Healthcare? LOL. Are we a leader in any industry or by any metric? I wonder if China is attempting to remove the last jewel in the USA's crown with these AI releases.
4) It must be refreshing for companies to have access to a new model that isn't going to get pulled because the government bans it 2 days after release. And it's open-weight so it wont go away--amazing--what a shift in the market.
5) If it becomes clear that open-weight models are the future of AI, will that pop a huge bubble in the US economy? Maybe. But, on the other hand, these companies aren't just training AIs, they are also building data centers which will remain valuable no matter what happens.
This comment is stunningly out of touch. The US dominates finance (overwhelming lead in assets managed, stock market capitalization, trading volume, and basically every other metric); space launch (about 85% of all mass placed into orbit); medical devices (exports nearly 2x the second largest exporter); and pharmaceutical research (more active clinical trials than any other country, near-universally ranked as the top producer of major pharmaceutical innovations, world-leading survival rates on many cancer types). We are the world's leading producer of both oil and natural gas. We manufacture the world's most advanced jet engines, precision agricultural machinery, and measurement instruments. We design the world's most performant CPU's and GPU's. We film the most popular movies and record the most popular music. Despite having only 4% of the world's population, there is literally not a single major industry where the US is not a leading player.
There is no doubt that China is a serious threat to US global leadership, and they have already surpassed us in many areas. But that doesn't mean the US is behind overall. We still have a lot of strengths and we have every reason to hope for continued success in the 21st century.
> (exports nearly 2x the second largest exporter)
Can you provide the source? In the WTO's broader medical goods category, Germany actually exported slightly more than the US in 2022 ($202.6 billion versus $189.6 billion), so such a huge difference in a few years?
> and pharmaceutical research
China accounted for approximately 44.2% of the 104 new molecules in 2025, compared with 26.9% for the U.S. and 15.4% for Europe.
The 2024 shares were approximately 34.6% for China, 30.9% for the U.S., and 22.2% for Europe.
> there is literally not a single major industry where the US is not a leading player.
I will just give 3 examples: China completed roughly 91% of the world’s shipbuilding tonnage in 2025, holds more than 80% of solar module manufacturing capacity and produces more than three quarters of global batteries.
Sure, these are my sources. World Bank (2023): https://wits.worldbank.org/trade/comtrade/en/country/ALL/yea... Observatory of Economic Complexity (2024): https://oec.world/en/profile/hs/medical-instruments
> China accounted for approximately 44.2% of the 104 new molecules in 2025, compared with 26.9% for the U.S. and 15.4% for Europe.
Source for this? My source was Citeline's 2026 annual review at https://pulseforinnovation.org/by-the-numbers-citelines-rd-a... which states:
> Findings from Citeline’s latest report, Pharma R&D Annual Review 2026, highlight the continued U.S. leadership in the life sciences ecosystem... The U.S. leads all other countries, currently advancing over 11,600 medicines – more than half of medicines in development worldwide.
> I will just give 3 examples
In general, I was referring to "major industries" in the sense of economic categories, as defined by (for example) the UN: https://unstats.un.org/unsd/publication/seriesm/seriesm_4rev... Here shipbuilding is considered to be part of "manufacture of other transport equipment." I admit my wording was a bit imprecise; if you are going down to the level of specific products like ships or solar panels, clearly it will be easy to find examples of things that the US does not produce on any significant scale. Nevertheless, a few comments on your examples:
> China completed roughly 91% of the world’s shipbuilding tonnage in 2025
Source on this? 91% is a lot and I can't find anyone else making this claim. However I'll concede that the US is far behind China, South Korea, and Japan in shipbuilding.
> [China] holds more than 80% of solar module manufacturing capacity
You are correct that the US has "only" 60 GW of domestic solar module production capacity according to https://www.utilitydive.com/news/onshored-solar-supply-chain... However, as context - the only purpose of producing solar modules is to generate solar power, and the US was the 2nd-largest producer of solar power in 2025 according to https://en.wikipedia.org/wiki/Solar_power_by_country Surely that makes us a leading player.
> [China] produces more than three quarters of global batteries
True but the US was the 2nd largest producer. In battery cells, we have an estimated capacity of 96 GWh in 2026. https://poweralliance.org/2026/03/18/american-energy-storage... Likewise in lithium-ion batteries: https://elements.visualcapitalist.com/ranked-the-top-lithium...
People are paying for tokens, that will likely continue even if open weight wins.
“State of [T]he Art” versus “State of [t]he Art”.
If not SotA then at least SOTA, which is more accurate.
In a world of so many acronyms, details matter.
I'm sure that more than once we've all had something we wrote which was intended to be obviously taken in a satirical or intentionally nonsensical manner taken seriously by someone else on the internet.
I’d be amenable to it but I’m already aware of them lying about token throughout by 50-100% on AA and on Twitter, even when you get a dedicated server.
Just too much lying piled up to extend courtesy here, even though you’re my favorite open model provider. Please, get your house in order, and if the house is so big that marketing drove this set, ask them to slow down a bit and market things that people can actually use / you actually provide.
Moonshot has committed to release it end of the month.
My only gripe was that it takes a very long time to think. It's a good model otherwise.