I know we have strong views on what a truly open model is (open weights, open training data, open training code etc.) but I really like how transparent they’ve been about the training of this model.
The realtime dashboard they shared during training (https://mimo.xiaomi.com/rl/) was an incredible learning and teaching tool for me, and they’ve been unusually comprehensive in sharing details about their methodology (check out that tech report - it's got lots of clever behind the scene tricks like Google or Deepseek writeups) and benchmark scores (even the stuff they didn’t do well on).
If you’re releasing an open model going forward, please consider offering the community more of this transparency!
I was absolutely mind blown when I saw how they were publishing that training dashboard while US models publish 100s of pages of reports (just provide a "copy as MD" button, folks, in the future). I was thinking about doing something similar but did not know how to show it, and this is a perfect example for someone who wants to show whatever they are training, for me it was local training on a consumer GPU.
My dream is to see this like a dashboard for a model trained across distributed machines, like Bitcoin mining, where minted coins are given to people whose machines were used for training. I don't know if they are worth it, but bragging rights alone, like a tag they can put on a website or social media, will be good enough for me.
For (C) -- I think it could additional create jobs funded by government, philanthropic and other private institutions. For example, a government funded museum may already be participating in Wikimedia GLAM projects (e.g. uploading historical images to Wikimedia Commons with complete metadata). Perhaps this type of open source contribution may increase if organisations realise their mission can be better accomplished by contributing this same open data into LLMs, in addition to Wikimedia Commons. If the museum's mission is to educate the public on the history of life in ACMEville, having LLMs be able to provide historical information and images to a prompt of "What is the history of ACMEville?" may be a good pursuit.
I'm sceptical though whether use of LLMs would encourage creation of data that doesn't already exist. For example, if you ask an LLM "What are the top 100 most prevalent flora endemic to ACME National Park", this data may not currently exist _at all_, and to collect, would require paying botanists to do an extensive field survey. If no one has done this work yet--why? Is it relevant to the scientific community, to making government decisions, etc, or just an obscure academic curiosity. There are certainly some journal articles on _other_ national parks describing some of their common endemic flora, but perhaps there was a reason for this. Such as a scientist funded by a one-off government program trying to determine how to preserve or even create habitat for a specific endangered species.
Consider for the prompt of: "What are the top 100 most prevalent flora endemic to ACME National Park"
An LLM may reply: "I couldn't find any journal article or other prior work that may answer this question. Typically such survey field work may cost $X to complete, require expert botanists, and take 6-12 months to complete. Let me know if you want further information on how to find and select a company to conduct such a botanical field survey."
Would this type of LLM response grow the industry of botanical field surveys, or do nothing, perhaps because anyone likely to fund botanical field surveys is already doing so regardless of whatever is happening with AI.
I think all of these can be addressed by being able to understand which machines contributed what to the model. I know that's messy, but if I can state "this information has come from Alice and the consensus is that it's good and right as it aligns with other information from Bob and others, meanwhile this information came from Mallory and stands out as being incongruent with the rest of the information I have in the model", we can identify malicious violations. If we're then able to state either Alice or Bob is one of the copyright holders, and one of them gets paid a little more, the other a little less as the confirmation agent, well, the economics of all of this changes a little.
At the moment we have Annas Archive being paid by frontier labs and rare/second hand books being destroyed in order to support the training regime. If instead we could just pay the publishers and they could distribute royalties to authors...
This dashboard is almost certainly built on verl (https://github.com/verl-project/verl), which comes with a bunch of dashboarding capabilities built in (that doesn't look too dissimiliar to these dashboards).
Not an expert on this but I think the RL runs need to work sequentially? I wonder what the opportunities for distributed execution would be... Maybe parallelizing the benchmark task or inference
Thanks so much for sharing this. As someone who mostly watches from the sideline, can you share what you can see in this dashboard that someone like me can't see? Is it the metrics themselves that they measure (the metrics tab is absurdly detailed), something in the notices, or something else I missed?
I might turn this into a blogpost if folks are interested, but my god there is so much clever info in that dashboard.
Here is one really neat bit:
A cutting edge training idea (for agents, it's been used elsewhere for ages) is on-policy RL, basically, it's not enough to say "here is an end to end agentic sequence (including tool calls etc.) that is perfect" you want to say "here is a sequence you might actually have generated that turns out to be correct".
Basically, it's more training efficient to improve models with small tweaks to do more of the right thing they are already doing sometimes than from some perfect oracular "this is the way" answer.
(if you've ever tried to teach humans new skills, you’ve probably noticed this too!)
When you do that, you care about how far the model you are updating (improving) has deviated from the one being used to generate rollouts (agentic rollouts for hard problems can take hours with lots of tool calls, so you can't keep redeploying every slight improvement).
Lo and behold, the dashboard literally has:
partial/avg_staleness (likely the measure of how many micro iterations the "generate answers" model is behind the "improving based on the occasional right answer" model)
train_infer_diff/new_infer/kl (a more direct KL divergence based way of measuring how differently the two models generate tokens)
How cool is that?!
And don't get me started on the clever ideas hiding behind dynsam/avg@n ...
Hold on, isn't that just standard practice for post-training LLMs for agentic use? Give task, generate n rollouts, grade rollouts (either at termination or after each tool call)? Or is the difference that the rollouts are generated ahead of time and then graded? (Of course, then it's not really on-policy.)
The best thing they did is being open about all the setbacks they had to deal with. They logged every restart with a reason, talked about dropping a cyber dataset after it degraded coding benchmarks. Also published real time training loss, benchmark scores after every checkpoint and running cost estimates.
Really the only thing missing was dataset descriptions, the dashboard only had random IDs like "dataset-zrso". I guess it's their lawyers fault.
It'd be great to see a description of even just a subset of training datasets. It feels very much under-reported how much expense is worth investing in preparing and selecting training datasets versus just using masses of random quality unprepared training data. This dashboard appears to be good though in showing the limits quickly reached when throwing parameters and compute at the problem.
For example, if they were to train on Wikipedia dumps, do they consider every article to be the same quality across each language, or have they done more work beyond Wikipedia's own article quality ratings to make training decisions such as "Ignore cebwiki it's machine-generated spam" and "Treat dewiki articles with coordinates within Germany as being higher quality (weight it higher) than their equivalent enwiki articles".
And let's say one of the datasets is all the source code of packages in the Gentoo package repository. Not every software package is a good example of how to write code. You perhaps wouldn't want to train your LLM on 1990s era PHP web application source code as an example of how to write code in 2026. Instead, you'd possibly want to use such PHP web application source code as a negative training example of what _not_ to write. But when training an LLM to detect software bugs, maybe outdated PHP source code is good for training.
Similarly for translation, perhaps UN treaty documents translated into 4+ languages are good translation examples because of high accuracy needed, professional translators being used, and bigger budgets. However this training data would perhaps be a negative training example towards translating chat messages, movie subtitles, etc because it doesn't use everyday slang and could result in output of nonsense such as "Pending Your Excellency's response, please accept, Your Excellency, my sincere greetings." for a prompt asking to write a birthday card for a child.
Preparing training data and deciding how to best use it for training I assume would be the largest expense (cost of labour -- mostly expert labour too) and also the greatest opportunity in the future for LLMs to improve. It seems to me somewhat irrelevant if the dashboard indicates a compute expense of $1m or $5m if good training datasets (prepared by experts in their fields) cost $10m/y to maintain. For example, hiring expert software developers to tag 1000's of open source software packages according to their quality, on different metrics, such as human readability, performance optimisation with choice of algorithms, reasonable trade-off between coherence and coupling in the software architecture, currency with state of the art programming trends/preferred dependencies/operating system APIs, etc. And keeping that metadata continually updated rather than a rapidly obsolete once off tagging project completed in 2005.
Because the world is conditioned to distrust chinese models (pick your reason here), I believe this is critical for them in order to kill any arguments outside the actual merits. They probably spent a lot of time making this call and might pay off on the long run.
Whatever well-founded/or not distrust people have in Chinese models, this dashboard proves/shows nothing that can make them trust it more or less. It's like providing the journalctl logs of your HTTP server on your website and claim this proves NSA isn't listening or something.
I don't work with large language models, but out of curiosity: mimo's score on DeepSWE keeps going up, so why are they stopping training early? Is it due to budget constraints?
It looks like the curve is flattering, and the current state actually looks slightly cherry picked (it matches a previous spike that looks a bit of an outlier before the result went down). The longer they train, the more they risk getting scooped by another release by someone else. Etc etc it's a judgement call based on all of these factors (and more, including cost/occupying a big cluster as you mention)
+1 at some point, you need to expect to train a much better base model using everything you've learnt. At the least, you probably want to bring on line the next 10 clever RL environments and ideas your team has been cooking up (which will pipeline into v2.7 etc.)
I wonder if this waters down the “distillation attack” claims by Anthropic. They have their own RL environments! I guess the caveat is that the RL datasets are still opaque, nothing is really proved.
> got lots of clever behind the scene tricks like Google or Deepseek writeups) and benchmark scores
Xiaomi MiMo is led by Luo Fuli, a former Alibaba & DeepSeek employee. Perhaps it is due to Luo just how similar Xiaomi's tech & GTM approach is to DeepSeek's.
maybe this is why Dario want to slow down AI development and all the big AI labs in the USA is singing the same song.
whey they all singing the same tune. it make me question what is their real motives.
they are afraid of Chinese good enough LLM model killing their margin. we already have story about US companies switch some task to use cheaper Chinese model hosted on Neoclouds.
The reason is money. They want regulation to make it harder for new competitors and competitors from other countries.
They invested billions into training the models but there is no competitive advantage, we see that within a couple of months everyone catches up. There is no way to profitability unless they get some policies to shields them against competitors that can't comply with the regulatory requirements.
That is also why there are things like Claude, Codex and Cursor. They are trying hard to build a customer relationship with a higher switching cost that hopefully sticks.
But the problem is that the AI buildout has become a large percentage of GDP. So obviously the government wants to keep it going because these companies are pumping enormous amounts of money into the economy.
> But the problem is that the AI buildout has become a large percentage of GDP. So obviously the government wants to keep it going because these companies are pumping enormous amounts of money into the economy.
They are pumping enormous amounts of money into each other. Hardly any of that is making its way to people, it's all going to highly automated construction and to energy use.
Seriously, how many jobs did the $1t in venture capital fund?
And, in this case, the dollar-amount increase in GDP serves as a virtual quantitative proxy for the increase in mowed lawns (and the value thereof). In other words, the participants in this economy are collectively ~$200 richer with their mowed lawns than they were without them.
If everyone goes around mowing lawns for each other, the economy is richer in lawn mowing at the expense of all the other things that would have been funded had everyone mowed their own lawns and purchased different services instead.
I am confused with this, if "everyone mowed their own lawns" then the net result will be exactly the same, everyone will be busy the same and not poorer, just without money movement.
This is not the same. If everyone wants mowed lawns, and everyone is busy working on that, there is no opportunity cost, everyone is working on their top priorities. The broken window fallacy is a fallacy because the headline gdp figure doesn't account for the destruction of the window which cancels out the benefit. In the grass mowing analogy nothing has been destroyed, useful and priority work has been done all around.
If the pricing is fair and at arms' length. What's happening in reality is as if they are mowing each others' lawns at wink wink nudge nudge $1000. Not a good proxy for actual value created.
In the real world, you have to pay taxes. So people are incentivized to claim less value for the lawns mowed, or even just do it themselves, instead of benefiting from the division of labour.
Leverage doesn't work that way. (If it were so easy, it would load up my investment portfolio with a lot more leverage than I currently do. And I don't live in the US where regulation T would keep me to a puny 2x leverage.)
But also importantly the government of the residents' country is about 39% ($78) richer, if say the participants are honest in reporting this and the country is the UK and the participants are people like you and me in the tech industry who frequent HN and would think to do something like this.
> Hardly any of that is making its way to people, it's all going to highly automated construction and to energy use.
How do we know that? How automated is the construction really?
In any case, the Fed and other central banks can print as much money as they want in order to hit any aggregate spending or inflation target they have for the economy.
I’m still not at all sure about the “billions” invested claim. How much of that is cloud running the models? How much is pre and post training (which may or may not be part of what we’d want to include in accounting). Etc. Does anyone have links to good reporting about this: not blind recitations of numbers, but analysis and thought mixes with investigation?
Well, if you are right, I just hope their protectionism will only affect the American market, and they leave us unAmericans free to get our models from wherever.
On the other hand (OTOH), China is desperate to keep up and keeps pushing open models (rightfully so), as they understand how far ahead from everyone the US is, and that whoever gets this right first basically is going to become an alien compared to others.
But even with all the open models the US is just insanely ahead in AI buildout and capital allocation (as usual).
Is it a bubble? Is it like the race for the-first-to-the-nuclear bomb? Both?
Unless something has shifted, “everyone catches up” is because these bleeding edge models are distilled. You don’t see this happening with other European and US labs and the problem isn’t something being ignored. I’m not convinced this pattern will continue indefinitely.
I did no such moral claim. I just noted that the foundation labs are working on technical hurdles to thwart distillation efforts and the cost and quality of Chinese models isn’t likely to keep up with the 6 month lag time everyone has assumed.
I don’t follow. Fair use is a copyright defense, and nobody is suggesting distillation attacks are just a copyright violation are they?
Aren’t they alleging these other companies directly entered into a contract and violated the terms, and in cases where question, answer pairs were obtained without such agreement, it was accomplished by outright wire fraud or theft?
I don't understand on what you are basing this reasoning? If a well educated workforce can produce Fable then why couldn't a well educated workforce produce MiMo?
Besides, the latter actually published and open-sourced its RL stack to make it reproducible, which would in fact make it more trustworthy than the models you are speculating were distilled.
No chinese lab has caught up yet. They've tried to fake it by distilling and overfitting on benchmarks to make their models look better than they are, the 'best' models available from chinese labs right now (GLM 5.3 and Kimi K3) fall apart completely when you try to do real work with them. K3 is especially embarrassing because it is larger than Mythos yet performs worse than opus 5 and 5.6 sol in benchmarks they haven't been able to fake yet.
Please explain how putting an upper bound on how good the strongest models can be prevents cheaper less strong models from catching up, rather than enabling it. I do not understand this argument at all.
The general idea is that Anthropic/OpenAI is pushing this narrative as an attempt at "Regulatory Capture"[1] which would allow them to make it prohibitively expensive for anyone but them to enter the market thus stifling competition.
You target the US companies: if they can't use these Chinese models, then they're less of a danger for a now captive audience in the US (and the West generally).
This is already kind of the case: the big enterprises don't really want to touch the latest Chinese models. It's a real pain, personally, I want to use them at work!
Show them you can burn tokens in seven sessions day and night with comparable results to Opus with less energy and less than 10 dollars a day, per dev.
We have. Unfortunately there are political realities that get in the way, and Bedrock for example doesn't have GLM 5.3 (Flash or otherwise) or anything new/useful
If it's hosted, all they know is "data goes to China".
Until profitable, reputable third parties host open models in the US with ZDR or they become plug-and-play for self-hosting at a modest cost, paying the US models is as much about data protection and liability as performance.
Yes, but thats not something a company engaged in regulatory capture for themselves care about: especially if they're worried they'll be outpaced and overtaken by the Chinese labs. Which they will be, IMO.
they care because they know it unlikely open weights will be banned, and thus available to American companies, with regulatory capture (onerous requirements) being a "good enough" "ban" that their big models don't face real competition, regardless of the open weight origin. American companies make open weights too, they are equally threatening to Big Ai financials.
...because everyone saw how well that worked for the Jones act, what with all the naval yards the US has lost over time, and how nearly no US-built ships operate where not legally mandated /s
it wouldn't slow down China as much as make it impossible for American companies to use non-American options, they care about their margins and don't want to be commoditized
I heard someone analogize token vendors to car manufacturers, where American companies only want to produce expensive options, the people want cheaper/better alternatives, and we ban BYD because those with enough money are more "persuasive"
The analogy is a good one, but your explanation is missing one aspect: the country (USA) does have a reasonable interest in having the capacity to build their own models. The “we need to slow down because it’s getting too dangerous” part is probably more related to “we need to slow our public facing development down so the US government can get the best and the American corporations can trickle out what we decide is safe”
It’s similar with cars. It’s not that American cars are better than Chinese cars on any tangible measurement. But America already shipped most of its manufacturing overseas. Everyone who built those factories is retired. The US should probably hold on to some capacity to make cars, seeing as their entire infrastructure depends on them.
American Ai/Car manufacturers could build cheaper/open models, some do, the big ones do not. It's not an either or, but a spectrum where they have chosen to build only in a subrange
That doesn’t explain the decline of German automotive industry which is now taken over by Chinese cars thanks to massive subsidies by the Chinese government
It’s not a reasonable comparison. In China, every corporation is de facto state run. The Party is in the boardroom and the executives are members of the Party. The CCP will build entire mega cities or pump money into this industry or that according to their plan. China isn't a "state-run" economy but a conditional-autonomy one. Private firms operate freely until they collide with Party priorities. Then the state wins decisively and without due process.
I'm honestly not sure why this is seen so negatively. It seems to be working pretty well for them, perhaps we should do similar instead of whining about others being more effective?
Another point of comparison we might make, how close is Trump's desired increase to the US Defense budget to what China is spending on industrial subsidies? It looks relatively close to numbers in these research papers.
> Please explain how putting an upper bound on how good the strongest models can be prevents cheaper less strong models from catching up, rather than enabling it. I do not understand this argument at all.
They are not proposing to regulate only the strongest models. They are proposing to regulate all models. If they are already on top, regulation may stop them from proceeding further, but it also stops the cheaper alternatives from catching up.
If they feel they have reached the asymptote of the curve, then regulation doesn't affect them, it affects those who have yet to reach the asymptote.
Particularly, the route they seem to want to go is "safety".
My guess is that Anthropic and OpenAI will push for "safety" regulations which require byzantine testing that, shocker, Anthropic and OpenAI can pass but the chinese models cannot. The route they'll take is import bans and potentially even general bans on products producing or using "unsafe" models.
They'll further likely try and push AI "safety" treaties from the US to other nations to further lock in their lead.
That's why, IMO, we've been seeing so many "OMG, AI will destroy the world and these AI researchers are so scared" articles.
Cracking down on proliferation of open models which can't be locked down using the kind of guardrails that Anthropic/OpenAI/etc insist are keeping the public safe from all manner of nefarious bioweapons, hacker swarms, propaganda bots, etc. They've discovered they can't meaningfully slow Chinese model progress, so the next best option is to knock them out of competition in the enterprise market for any American company.
Both Anthropic and OpenAI leaders have repeatedly made this exact argument that it's impossible for open models to rigorously enforce the same kind of safety framework as proprietary cloud-served models. It's implicitly part of any regulatory framework they advocate or else it wouldn't be "fair" to American companies since Chinese models would "cheat" (provide weights).
This is true, and it's a good point. I agree that open weight model regulation would either limit the intelligence of open weight models below the frontier or kill them entirely. It would not prevent closed weight Chinese models, but those aren't really tenable in American enterprises unless they strike deals with American cloud providers to deploy them, via products like Bedrock. I unfortunately also have seen no evidence that we can put any sort of guardrails on open weight models whatsoever and so am reluctantly convinced that they should be regulated below the frontier until such a time as someone comes up with a mechanism that is not circumventable.
This is definitely part of it. I think the reports/PR over the past month ended up being a serious unforced error.
Chinese models are increasingly closer to the frontier, while being able to run on much cheaper hardware than what US frontier models run on.
On top of that, both Anthropic and OpenAI showed that they can't really be trusted on data security.
Even if US companies can be forced to not use Chinese models, the rest of the world is going to see the risks and the availability of good enough open weight models for their purposes and be more likely to lean in favor of self-hosted Chinese models or local inference clouds.
> and the availability of good enough open weight models for their purposes
it is childish to believe Chinese are going to give up profits to provide such open weight models forever. the whole idea of being "open" is not compatible with the Chinese culture.
This almost racist read of other cultures has always seemed so bizarre to me. Even moreso when said as an argument on the side of completely closed competitors, some of which are outright seeking to ban open weights.
Chinese companies will continue to provide open weight models as long as it is profitable to do so. Chinese companies are on the more open end in many other industries despite the lack of meaningful foreign competition (for one, 3d printing) so there's plenty of reason to be optimistic as far as I'm concerned.
In a recent Dwarkesh podcast Dylan Patel breaks down how little compute the chinese labs actually have- not even the fact that they don't have access to new Nvidia chips and they're stealing them through shell companies- just that, even if they have cheap electricity, the compute just doesn't compare. Maybe even two orders of magnitude less. They couldn't get it even if they had the money. And if you look at how much more efficient newer chips are, that cuts the effective compute in half again. The conclusion was that they are at least 2-3 years behind.
For frontier labs the current compute seems to be driving model progress (in training) at least to some degree, even without true RSI, and this seems like it'll continue to keep any chinese model from drawing even with the frontier labs, at least for the foreseeable future.
Inevitably the chinese government will drive more funding in chip fab technology and the money will come around to build chinese data centers, but who knows how far off that is. A few different things in the tech tree need to fall into place. It doesn't seem like it'll be next year.
The counterpoint to that, though, is that the Chinese companies have to figure out how to be competitive, regardless of their significant compute deficit. And, as far as I can tell, they're actually doing that. They're trailing the frontiers in model effectiveness, but not by years. It's single digit months.
If there is no upper bound how how these things scale with compute, and if China does really begin to catch up to Nvidia (and they're probably not going to feel encumbered by US patents for domestic AI hardware, given how important AI seems to be to the Chinese government), there will come a day when China leapfrogs the US on AI.
Does anyone know what are the proposed regulations? Controlling software is impossible, so the only option is banning hardware ownership. No more mac studio.
If you pay attention to how these US CEOs talk, it'll be "safety". If I were to guess, they'll try and require a lot of testing, validation, certification before a model is legally allowed to be used in the US or on US products.
It won't be a great moat, they'll probably try and get trade treaties setup to try and expand the moat. But ultimately it won't slow down chinese model development, just limit who can legally use them.
From the frontier labs, the only publicly stated one seemed to be to give them an exception from anti-trust laws to form a cartel and place - incidentally friendly - regulators in charge of monitoring everyone's work.
From politicians like Bernie Sanders, we've had proposals like 20 year imprisonment for anyone researching "ASI".
OAI and Anthropic are forced to release a better model every x months otherwise the Chinese ones will not only be cheaper but also better.
So how could Dario show the investors very nice profit charts representing profit = revenue excluding training costs if it needs to pay a lot of training every x months?
They want to sell the same model for longer(a kind of software subscription where the cost of running /inference is cheap) but the Chinese don’t let them do it. That’s the gist of it. You can see already how they nerf the models just a week or so after release and try all kind of tricks to deliver you shitty performance for the same money. I think it’s part of the same issue of costs and enshitification plan.
In the meantime let’s hope they don’t get to ban the Chinese models(I think they won’t), local AI hardware will get cheaper and the whole AI doom saga will slowly fade to the point that Anthropic becomes a kind of IBM stuff with proprietary data, enterprise certified alignment and enterprise contacts. Think of Accenture junk.
Competition on the provider side—when no single dragon monopolizes the sky—brings fortune for all.
As normal consumers with common sense, we should never naively assume others care for the world out of the goodness of their hearts. Maybe they do, but we should never rely on that.
We can only get good, affordable deals when there is enough competition on the other side.
I mean DSv4.1 Flash and GLM 5.3 kept in check by a supervising frontier like Astra or Fable already in my experience clowns massively on ever using Opus or Sonnet. Opus 5 in particular has been such a stinker that they have to know that they're going to get smoked outside the halo models.
Dario has always wanted the AI development to slow down and be more careful. Safer AI development was a core reason that Anthropic split off from OpenAI.
What's different today is that now all the big LLM firms want to slow down AI development. When men like Musk and Altman (both known for habitually shooting their mouths off and saying whatever they need to whoever needs to hear it regardless of truth) suddenly agree with Amodei, that's when things start to smell off.
> What's different today is that now all the big LLM firms
not all, just a few American ones (~PayPal Mafia + Google), there are other big American LLM developers (notables include Nvidia, Meta, and Palantir) that do not agree
The realtime dashboard they shared during training (https://mimo.xiaomi.com/rl/) was an incredible learning and teaching tool for me, and they’ve been unusually comprehensive in sharing details about their methodology (check out that tech report - it's got lots of clever behind the scene tricks like Google or Deepseek writeups) and benchmark scores (even the stuff they didn’t do well on).
If you’re releasing an open model going forward, please consider offering the community more of this transparency!