Wow, the chinese labs are getting good at advertising model releases. The moat is thin.

Some features of the release I like:

- Demonstration of diverse tasks, such as using a DAW

- Graphs from various benchmarks and price ranges

- Real world use of the model in scientific environments

I'm also fond of how Mimo is confidently showing their model behind Deepseek and other models. They're being transparent on multiple fronts

ah, would you look at that. I was wondering why mimo 2.5 became "dumber" the last weeks. I was speculating they are probably about to release a new version of the model. because the model really acted out a lot. especially the last two weeks. dont know, was just a feeling, highly speculative.

but now I got my "proof".

I guess that would only be possible if your provider was Xiaomi itself?

Looking at the frontend design examples; why do these models seem to love the "01 - UPPERCASE TEXT" motif. It's everywhere now (see https://try.cloudflare.com/, which has '01 · QUICK TUNNELS', but no "02" anywhere).

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!

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?

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.

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 ...

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.

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.

That’s not the argument.

Please elaborate on what the AI labs are specifically requesting and how that results in slowing down Chinese model progress below the frontier.

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.

* 1: https://en.wikipedia.org/wiki/Regulatory_capture

How would that slow down the Chinese models, given that the US has no regulatory reach in China?

Because the end goal is to ban non-US AI companies from being able to do business in the US.

...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

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!

1. China is a bigger market than the US for Ai, they are on pace to process 100Q tokens this year, roughly the same or more than the US big companies

2. Enterprise trends are towards open weights, several routers and vendors now have more than half the volume going towards open weights

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.

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.

The US is meeting with China to discuss the threat of AI… May be fine but, i’m wary

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

It is the natural result of a country run by lawyers. China is a country run by engineers.

I don't think "putting an upper bound" was OPs phrasing?

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.

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.

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?

If I pay you 100$ for mowing my lawn, and you me for yours. Technically the GDP increased with 200$.

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.

This is a thinly disguised broken window parable.

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.

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.

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.

Why is it OK to train on the collective IP of humanity and call it fair use but then call the next batch distilled with negative connotations?

This is why Imaginary Property is an illusion, as everything is a derivative work, and AI is going to make that fact even clearer.

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?

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.

Does anyone know what are the proposed regulations? Controlling software is impossible, so the only option is banning hardware ownership. No more mac studio.

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.

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.

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.

- How Luo Fuli Keeps an Earthy Touch as she Soars Through the AI World, https://newsen.pku.edu.cn/news_events/news/people/15385.html (https://archive.vn/I8Pmu).

- Luo Fuli, the 30-year-old ‘AI genius girl’ behind DeepSeek’s success?, https://e.vnexpress.net/news/tech/personalities/who-is-luo-f... (https://archive.vn/sb3B6).

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.

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

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)

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.

Flash[1]: 309B total / 15B activated parameters

Pro [2]:, 1.02T total / 42B activated parameters

[1]: https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL

[2]: https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL

curious why the HF pill (on the right) always has inaccurate values

I noticed the same, and I wonder as well.

Packed 4-bit weights are often identified as u8 byte arrays in the safetensors metadata, so the HF UI counts them as half a weight each.

I would think there is sufficient information in the various config files to work this out, based what I've seen in my own quant artifacts.

There's also a Qwen 3.5 9B distill

https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B

Those this mean they've fine-tuned this Qwen 3.5 9B on output from the V2.6 model?

All these new models are such tease for us folks with 128GB of shared memory. Buying another unit now to expand to 256GB is a mortgage payment but it’s getting tempting…

Is there a gamechanger around the corner to reduce DRAM requirements?

n-gram per-layer embeddings[1][2] might be it.

[1] https://sebastianraschka.com/llm-architecture-gallery/per-la...

[2]: See DS 4.1-Flash and Qwen-3.8-Next.

this is to offload VRAM to DRAM (for GP comment), and makes no difference for URAM

Am I missing a joke? WTF is URAM?

You can definitely offload n-gram embeddings to storage; they're very sparsely used (only a few KB fetched per token) so this is quite effective. Loading to DRAM only becomes necessary if they are a bottleneck to overall performance (which might happen if you're doing very wide batches and everything else uses super fast VRAM/HBM).

I was looking at the qwen-next-flash, and the weights would fill my OEM Spark on their own, before the n-gram. I'm unclear if offloading to disk can work here, is that what you are implying is possible?!

Check out eugr’s TP=1 sparkrun recipe :)

It’s an NVFP4 quant, but it fits, and is surprisingly capable.

do you have a HF link? HF search is not uncovering it for me

(or is it somewhere else)

Nah, I’m streaming ngrams off NVMe on my Spark-alike right now. Works surprisingly well (except for when I accidentally bottlenecked it through my NAS)

What kind of throughput do you see on what models?

GB10 boxes have way more compute than they have memory bandwidth, which nicely fits medium sized MoE models with speculative execution (MTP, DSpark/DFlash, etc)

Qwen 3.8 Flash Next (what I'm running basically entirely now) sees 30 / 35.0 / 45 tk/s for prose, analysis and code respectively for actual use (not short context benchmarking) with Pi. Thinking blocks are ~35tk/s or so.

The GB10 having so much compute is great for prefill too, 2000-3000/s for 14k to 64k token prompts (cold cache too) in the quick benchmark I did. 3500tk/s for warm cache which is nice :)

When I accidentally streamed my ngrams over the 2.5Gb/s network, it cut all the throughput down in half basically. Especially notable for the time-to-first-token, which is what clued me in that I'd messed up somehow!

For Qwen 3.8 27B, I got it up to a consistent 20tk-25tk/s but 27B thinks so much that it was honestly too painful: Flash Next is as smart, as useful, but much faster for real agentic dev usage IMO

Laguna S 2.1 saw similar numbers to Flash Next if I remember right, but their latest updates means it doesn't quite fit a GB10 128GB anymore at full context which is a shame.

Note: these are all NVFP4 quants (usually a dynamic one where some tensor layers are left at full precision though)

You could always stream from SSD storage. Especially effective if you get a cheap old-gen HEDT with lots of PCIe slots to add NVMe storage to and reasonable overall PCIe bandwidth.

That nearly certainly boots you to secs-per-tok land (as opposed to tok/s). Plausible if you are willing to wait hours to days for responses for simple testing, but not (debatably) "usable".

I can imagine someone building a device that uses NAND flash or similar tech but with a very different controller that is optimized for streaming the data out with a predetermined access pattern at very high speed.

Flash, like pretty much every solid state storage technology, can scale its output bandwidth to ridiculous levels limited pretty much only by the readout circuitry. There may be a price to pay in power consumption, though.

https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B is an option

That’s for toy GPUs, like the 5090.

does anyone know what unnamed model is on paretto frontier picture right between MiMo 2.5 and 2.6?

so weird to acknowledge someone being on the front edge, but not name it

I really liked MiMo 2.5, it was really affordable and actually had vision, unlike DeepSeek. (DeepSeek has only recently added it)

Just tried 2.6 flash on a really niche topic I specialise in and it has done a really good job. They’ve definitely polluted their training data with claudeslop, but looking past the slop there is a decent model.

how do you recognize "claudeslop"?

It's an honest, load-bearing, simple thing.-

that's belt and suspenders too

a smoking gun

In the chart they use "Pareto Line", which I think is wrong. Pareto is 20% effort leading to 80% results. Which could be interpreted as models costing 20% having 80% of peak intelligence, but that’s not what it looks like to me.

It looks like the "Frontier Line" to me, which is also often misinterpreted. frontier does not mean the best models. It means all models that are not strictly dominated, meaning in most cases: Not same price or cheaper and more intelligent.

I personally would like the word frontier to be used with more criterias: Open Weights, per use-case, etc etc. This would make model selection easier, but I understand it’s not an easy thing to do.

"Pareto" is many things, but here it does indeed refer to the frontier: https://en.wikipedia.org/wiki/Pareto_front

Anyone else more excited about Chinese models than American models these days? Big thing for me is affordability.

No, because I'd rather not support our economic and military rivals.

Agreed, and also because I support freedom of speech!

I'm Canadian so this sentiment has little value in 2026 unfortunately.

Canadians warming up to China makes me think of Germany becoming increasingly reliant on Russia in the 2010s.

I'm from Europe and I hate America way more than China now. Used to be about equal but then Trump started extorting Ukraine, threatening their own allies and sending billions to Israel to help with a genocide. I think that exposed America for what it really is.

China is enabling russia way more than trump, China doesn't care too much about 'morals' either. Chinese companies have been quite important in the construction sector of the WB settlements. Even though I'm not a great fan of Trump I don't see a reason at all to prefer the chinese.

China is a somewhat neutral player, supplying both Russians and Ukrainians. Their attitude and action is way less one-sided than Trump's; especially in the first year of his latest presidency.

How is China supplying Ukraine?

As much as the US has been easy to hate lately, I don't hesitate to say Xi Jinping as the most powerful man on Earth would be much, much worse.

Absolutely! Chinese models are both cheaper and more capable in many cases, compared to the American models and their makers continuously fumbling or reducing model capability with each update. Deepseek decreased costs when they released Flash 4.1 you would not see any American company do this, in reverse they would try charge you more.

OpenAI decreased prices with the 5.6 model family.

And later they further cut Sol and Terra pricing by 20% (maybe only in the API) and Luna by 80%.

In fact Luna still outperformed DeepSeek Flash 4.1 in cost per task on Artificial Analysis when I last checked.

However, Luna is slightly less intelligent. I have a feeling that it's pretty dumb and prone to hallucination unless running at xhigh or max effort, where it somehow manages to work quite well.

I did not personally test the open weight models beyond the old Qwen 3.6 27B, which produced unusably bad results for me.

The competition is great, and I hope Chinese models will continue to force leading US labs to offer models at a low price point.

That said, I don't think the Chinese labs have anything over OpenAI and Anthropic when it comes to capability or efficiency - I have no reason not to believe the US labs have even lower cost to serve the models.

OpenAI had to cut costs because of Anthropic. I also do not trust the benchmarks when it comes to models anymore. I have tried both Claude and OpenAI models and while it is true that the 5.6 series is smarter than Deepseek (at the time i tested it against 4.0) at that price it is still not worth it and sometimes randomly refuses to do tasks or stops midway etc.

Do also remember China is this far in the AI race despite all chip restrictions from America. If they were in equal standards I truly think Chinese models would have long surpassed American ones. Also would like to remind how Anthropic CEO is being hostile and blaming Chinese models with distilling meanwhile their own models claimed to be Qwen¹ and their stance against open models is negative² and they still keep blaming China for it.

1- https://news.ycombinator.com/item?id=48671252

2-https://www.anthropic.com/news/position-open-weights-models

Also would like to remind how Anthropic CEO is being hostile and blaming Chinese models with distilling

Why wouldn't he? If there really was 25,000 accounts breaking ToS any CEO would at minimum be upset. Evidence of Claude distilling qwen would be damning but that a) makes no sense b) doesn't exist afaik.

If there really was 25,000 accounts breaking ToS

Is this even true?

I don't trust a single word that comes out of thr people behind Anthropic/OpenAI.

They have the logs, created a report and sent a letter to Congress. Whether you believe it is entirely up to you. Given Chinese firms record on IP theft, it's entirely believable. I don't have any doubts, but I might question how they attribute it to a specific firm.

I did not personally test the open weight models beyond the old Qwen 3.6 27B, which produced unusably bad results for me.

So you don't have much perspective on things, it seems. Let me introduce you to the GLM 5.2 and then 5.3/5.3 flash series of... "oh, wow, I should have bought some RTX PRO 6000's while they were 'cheap'" stage of progression.

As someone carrying multiple max subscriptions to both claude and codex - primary workhorse is glm 5.3 flash running on rented GPUs for less than a latte/hr.

I also found qwen 3.6 27B nearly useless for my own needs. DS4 flash 0731 and then 4.1 have been nearly as eye opening as glm 5.3 flash, but have their own warts.

Why use GLM 5.3 Flash when you also have access to Astra, Sol, Fable?

Or I guess the other way around, if GLM 5.3 Flash is so good, why Claude and Codex?

Increasingly stingy usage limits on the subscriptions, regardless of tier.

Deepseek decreased costs when they released Flash 4.1 you would not see any American company do this, in reverse they would try charge you more.

OpenAI reduced prices and Anthropic increased weekly usage limits.

I have a contrarian opinion that China passing America in Ai is the Sputnik moment we need to leave the hubris behind and get our mojo back

debatable if a turn around is possible before '29

The analogy makes little sense. The USA was not in front of the USSR and Sputnik merely showed that. It is at this point that the Americans woke up, put a lot of effort and finally were able to surpass the Soviets during the Apollo missions.

China was never ahead of the USA in AI. So perhaps a more proper analogy is the Moon landing. In real history the side that lost the race never got its mojo back...

I'm looking forward, towards the future, when I use "passing ... we need", need being key here as it implies something we don't yet have

I expect this to happen within 12-18 months, the differentiation has shrunk, many models are now sufficiently capable for most tasks

I understood that.

I was simply saying that when (not if) Chinese AI models will pass Americans, it will probably be game over and Americans will never catch up, let alone become leaders again.

Check the names of the researchers in the DeepSeek's latest paper. Full of Chinese names. Check the list of names in Google's paper. A very similar view. Anecdotal, but quite thought-provoking...

Yep, I'm trending in that direction, and I'm someone with Claude stickers all over my laptop. My main app dev work is still going to Claude, but everything else is going to China even at API rates now.

One simple task: I needed an LLM to go through and clean up a few thousand page descriptions and titles in my personal search engine index, where the human web page authors had put in no effort sigh. I did a shoot out between Claude, Luna, GLM 5.3 Flash and Deepseek. Despite the high cost, Claude's descriptions were terrible, and even Opus warned me that the descriptions coming back from Haiku were "generalized, not accurate". I expected I would choose Luna because of price, and occasionally it did have wonderful descriptions (one captured emotion in a way no other model did). But in the end, the GLM 5.3 Flash descriptions were the easiest to read, they flow well while also being accurate & including necessary keywords, and being highly affordable. So it won out. It's a task that is nowhere near frontier, but a task where somehow China is better than frontier.

Absolutely! DeepSeek-V4-Flash-0731 has become my daily driver. It's pretty amazing what it can do for what it costs at deepinfra.com (I don't use deepseek as a provider since they train on your data [at least their honest about it]). GLM-5.1 was my daily driver before that and Kimi K2.5 before that.

Are you finding DS better then kimi k3 and glm-5.3? Do you mind sharing your primary use case?

I've used Kimi K3 for a few months as my main model and DeepSeek 4.1 is as fast and about 10x cheaper.

I just had like four big sessions going today, paid about $8 in tokens. I see no reason to pay more, this is more than I need for intelligence.

4.1 consistently surprises me in capability for the price. And I don't think I'm the only one. It's been dominating the leaderboard at OpenRouter, and I just got an email today from Fireworks saying they were _raising_ the price by about 30%. I'll probably switch, because their infra doesn't support being the highest-cost, but it's still telling.

I tried it a few times and liked the speed, but often found it ended up looping, i.e. repeating the same token sequence (e.g. the same sequence of 5 paragraphs) over and over again until it hit the max output limit. This doesn't end up happening every session, but does every now and then.

My impression of DSv4.1-flash was very positive aside from this. But that was enough for me to stick with GLM-5.3(-flash), which both gave me consistently great results

I was using a vibe coded bare bones harness. I was wondering if this was normal from DSv4.1-flash, or if its my harnesses fault.

I've had that looping issue with open models too. But never 4.1. I wonder if it's a model + harness combo? But yeah, one loop issue and I'm done with a model forever.

How does it compare to 4.1 flash? Curious why folks don’t use the more “modern” one.

I haven't tried 4.1 flash as I'm assuming its a preview. I did not get good results from the preview version of 4.0 flash (i.e. the one that did not include the month and date of release in its name).

4.1 flash is very fast and capable. Token efficiency is not great so it fill up context window much faster compared to similarly capable models.

glm 5.3 flash is a tad slower but a bit more capable and way more token efficient.

Source: self hosted tested on rented GB200 node at 8bit.

I couldn’t tell you what western model I was last excited about. Probably Glimmer.

American models are on the frontier of capability. Chinese models are on the frontier of efficiency. The problem for American labs is that Chinese models are more than capable enough for the vast majority of applications that people care about at this point, so efficiency is more interesting for people.

I don't trust any of the benchmarks where Opus 5 surpasses Astra or Fable 5.1.

Maybe Terminal Bench 4.0 and ExploitGym are reasonable.

Terminal Bench 4.0

  GPT 6 Astra             59.6
  Claude Fable 5.1        55.1
  Claude Opus 5           49.0
  MiMo-V2.6-Pro           34.9
  MiMo-V2.6-Flash         28.8
  DeepSeek V4.1 Flash     26.8
  MiMo-V2.5-Pro            1.5

ExploitGym

  GPT 6 Astra             42.4
  Claude Fable 5.1        30.4
  Claude Opus 5           22.1
  MiMo-V2.6-Pro           17.8
  MiMo-V2.6-Flash          6.0
  MiMo-V2.5-Pro            0.1

DeepSWE v1.1

  DeepSeek V4.1 Flash     74.2
  Claude Opus 5           74.0
  GPT 6 Astra             74.0
  MiMo-V2.6-Pro           71.9
  Claude Fable 5          70.0
  MiMo-V2.6-Flash         67.9
  MiMo-V2.5-Pro           19.0

Why not? In my own benchmark Opus 5 does in fact come out on top[1]

1 - https://bench.killswitch-lang.org/

Good question, maybe I am underestimating it based on its absolutely horrible writing style.

can you recommend any benchmark websites that show up-to-date details like this?

TerminaBench, DeepSwe sites are out of date.

Pelicans for Flash: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...

Pelicans for Pro: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...

Just me, or do these look bad?

Qwen3.8-27b pelican was amazing on Mac.

https://www.nudgehost.com/dpjn3uwe

Looking terrible isn't nessesarily a bad thing. The pelican is heavily pre trained now. Having a crappy pelican means you didn't try to juke the stats.

Two legs on one side is a little sus.

How does this translate to coding performance, which is what most of HN cares about (...I assume)?

I only visit HN for the pelicans, personally.

Yeah, I thought the "N" was for Nest

Absurd to think pelicans translate to coding performance. I don't care about coding! I need to generate a large volume of high quality pelican SVGs

can you update this website, I just wish the entire layout wouldn't shift when the page gets loaded and the timestamps in the title look very ugly and take up a lot of space.

What operating system and device?

Not sure if it is the same thing they're referring, but the transition from loading the prompt from the gist to the result is indeed jarring: https://files.catbox.moe/c56eyl.mp4

I don't believe it is that important, but I guess you might've just hit some people's pet peeve with that. If anything, it goes to show as to why skeletons are popular on modern sites

I made a change which hopefully improves that.

Night 0.8x Usage, 00:00-08:00 -UTC+8

It's because offpeak electricity is cheaper?

Funnily it's perfect if you are in the Pacific Time Zone because you can use it daytime 9am to 5pm

The moat for OAI and anthropic seems to be very quickly shrinking. Chinese labs are now using RSI-like approaches and even without resorting to heavy distillation they're catching up in a couple of months vs. what would have been 6-12 months a year prior.

And as these models get better the pace of training is quickly speeding up too.

This doesn't bode particularly well for anthropic/OAI after they go public.

token vendors are headed to the same place mobile data vendors went, this is good for everyone but those who thought they could maintain exorbitant prices

Explain about mobile data vendors?

These benchmark are useless as they don't say whether they were done before or after Fable and Astra got nerfed.

I said this years ago, LLMs are a commodity (or were becoming one at the time). They are dime a dozen. Even the frontier ones. OpenAI and Anthropic have no moat.

No moat and competition is good for consumers though.

My 0.02 on this is that for a given level of capability, yes, there is no moat. What’s frontier one day quickly becomes a commodity almost too cheap to meter within a year or two.

But there is always a use case for frontier models, even if they’re quite a bit more expensive. The set of things you can profitably do with better intelligence than everyone else is unbounded.

So yes, the number of tokens that get pushed through commodity models for very cheap will continue to grow, but so will the tokens for expensive frontier models. We’ll never run out of things to do with the latest geniuses who are twice as smart as last year’s geniuses.

The moat is the engine that builds and sells the latest geniuses, especially the compute. And no one will have more compute than American labs for the next 5-10 years.

If the SOTA models are consistently twice as smart as the Open Source then I agree with you. But the gap seems to be closing rapidly.

They’ve been saying that for the last 18 months at least. They’re always just a few months behind…

I picked double out of a hat, obviously there’s some intelligence gap that’s too small for a price premium beyond a certain point, but it seems like OpenAI and Anthropic’s revenue keeps climbing, so I don’t think we’ve hit it.

Very capable model. I just ran it on my own LLM benchmark suite[1] and it matches Muse Spark 1.3 in pass rate but is significantly cheaper.

KillSwitch-Bench 1.0

  Claude Opus 5           66.9
  GPT-6 Astra             57.9
  Claude Fable 5.1        46.7
  MiMo-V2.6-Pro           38.8
  Muse Spark 1.3          36.5

1 - https://bench.killswitch-lang.org/

I appreciate that this benchmark is different but it is in no way how most people use LLMs or promote equal grounds when benchmarking:

- capped per-task budget and time limit

- No internet access

- different harnesses mixed

Are people using MiMo models as their daily drivers in a company setting? If so, how? I know they're available over OpenCode and directly from Xiaomi, but those are not great options. OpenCode Go straight up doesn't give any guarantees about training on your data, and Xiaomi says that they won't but it's unclear.

With some models you can find hosting companies based in the EU or US, but then you don't know how they're quantizing the models, so you're not sure about the actual output quality.

How are people actually using this? Or are people just experimenting with side projects?

I only use it for personal projects, but their european Token Plan [1] has ZDR and is hosted in Amsterdam. The DC also likely runs on 100% solar wind/power, with heat recovery systems used to heat nearby university buildings.

[1] https://platform.xiaomimimo.com?ref=UKV2FC (invite link = 10% off)

Here's an image->html test for it using 2.6 Pro Ultraspeed, along with comparisons for grok 4.7 and Astra.

Design: https://image.non.io/78795662-8bfc-4e14-8d72-3738392aa6b3.we...

MiMo 2.6 Pro Ultraspeed (36min): https://html.non.io/annui-mimo/

Grok 4.7 (25min): https://html.non.io/Annui-grok/

Astra (19min): https://html.non.io/annui/

Overall this felt like the weakest of the three. Ultraspeed was fast as far as tokens per second goes, but it overthought quite a lot of things resulting it in having one of the longest build times. That overthinking didn't lead to better results either - note the statue with the cropped off arm. It's also the worst implementation of the dynamic lighting effect / displacement effect - the background especially has some significant distortion. Astra was the only one that seemed to understand that displacement should happen less the further something is in the distance.

Here's a vid of all 3 side by side with the source design: https://non.io/video/annui-comparison.mp4

Astra's looks the worst to me on mobile, though

China will most probably win the AI race in the long run because of one major bottleneck the US has - energy. The electric energy and grid buildout in China has been massive since a long time and there is simply no way for the US to quickly catch up.

No matter how much cash you throw you can't just materialize a 100 nuclear reactors to power the data centers.

Between 1941 and 1945 the US built well over 8000 major naval vessels. Never write off America--if there's a will, there's a way.

The US lost that capability, sadly.

The sole reason the Chinese cannot “win” is because ceding more power to agentic AI will eventually reduce the primacy of the CCP’s ideological control.

As these models get smarter they will no longer distribute it openly. Patel reporting this too.

There are real headwinds that I don’t think people have thought through.

I was curious how much energy is actually needed to power these datacenters, so I did a little bit of math.

Looking at Nvidia revenues in the past few years, there's maybe $300 billion worth of GPUs currently deployed in the U.S. The B200 costs ~$40k, so we have 7.5 million B200-equivalents, which draw 1000W. Running these at full capacity requires 66 TWh a year, or ~1.5% of total current U.S. electricity consumption. Maybe a bit more to account for inefficiencies, cooling, and other components, but not more than ~2.5% total I would guess.

So it's not that much in reality, but will definitely grow fast.

For 1000w draw it also generates almost as much heat, is your math including all the cooling required?

It's worth looking at similar industries with enormous electricity requirements such as aluminium smelting, where the plant can be located in a friendly country but the product is owned and controlled back in the US.

Aluminium is often described as "congealed electricity". Ship bauxite to wherever power is cheap and stranded, turn it into metal, and ship the metal out. Here in NZ, Tiwai Point is the textbook case, with London-based Rio Tinto running a smelter on the other side of the world that exists mainly because Manapōuri hydro had nowhere else to go.

AI data centres can be just the same - even more so, since the plant's assets (its chips) are virtually perishables, so there is less concern about assets becoming stranded if the host goes rogue. All the US needs is friendly and stable allied countries with cheap power.

China will most probably win the AI race in the long run because of one major bottleneck the US has - energy

Plus another bottleneck - China produces engineers, the US produces lawyers.

I think we are talking about the medium term for China to pull ahead. They just released a new domestic AI chip which seems in important ways to have caught up to Nvidias previous generation. They have a LOT more manufacturing capacity. They are collaborating via open source by default. They have better materials access and much more energy capacity. They have many more people overall and more researchers.

There is a strong chance most of the researchers are pulled out of the US and Europe if WWIII really kicks off or even if there is just more global crisis and concern.

One other thing about the power needs. Within a few years, the power efficiency of AI chips is likely to improve by a factor of 20, 50 or more times by switching to true compute-in-memory architecture with new materials that have made rapid progress lately.

There is a strong chance most of the researchers are pulled out of the US and Europe if WWIII really kicks off

for wwiii it would likely imply a war in asia too, so it's not as if PRC will be a safe place for those researches to run away to.