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!
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.
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.
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
Chinese companies are well documented in their copying / following / training on output / outright stealing tech.
China, as far as I know, has not come up with anything on their own and has only been able to play catch up and the memory optimization game as far as innovation.
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 people with enough money are more "persuasive"
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.
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 ...
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.
> 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.
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.
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.
Apologies for not taking the time to find it, but there was a post that tried to determine if the pelican was benchmaxxed across a bunch of models by comparing it to other SVGs, and found that it wasn't at all.
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.
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.
> 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.
> 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.
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.
API rates still aren’t quite competitive with the OpenAI x20 accounts, but they are definitely getting close with deepseek 4.1 flash. I spent a few days with only 4.1 and was very impressed.
Also, frankly, as a fellow Canadian it's pretty clear that the biggest "rival" the US has right now is itself. Just passed out in the corner puking on itself shouting about all the foreigners who won't talk to it.
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.
Yes, someone can still blow up a pipe and they look the other way. On the other hand, you can also draw parallels to themselves becoming increasingly reliant on US vs UK in the past.
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).
My guess is that by function they break down frontend sections or components into pieces and I believe document things for themselves on some level, or purposely are verbose in this way. It is probably also shaped by users and existing web patterns. They probably get reinforced by models the more common they become.
The extraneous small-caps labels are one of the main idiosyncrasies of AI generated markup. I wonder how much of this is a "scaffolding" technique to help the model build stable designs. But was it reinforced in RLHF or an emergent behavior of the models?
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.
They match my experience. Astra and Fable I rate below Sonnet. They are incredibly poor. They were excellent for a couple of days after release and then plummeted.
Maybe I am being routed to more quantised versions or less capable models with system prompt to fake Astra or Fable.
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…
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".
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?!
I'd recommend pointing your agent at it (after installing sparkrun), and asking it to research the absolute latest in TP=1 Flash-Next - mine grabbed particular vLLM nightlies and mods to improve performance, and it was well worth it.
I have a quirky vLLM on k8s on 2x OEM sparks setup with 9 models available to me. I'm not keen to run nightly vLLM, too many issues with it in the past. Going the qwen-next path means displacing things I use daily :/
I have a watchful eye on the diffusion ~ Jev/Kev PR
For what it's worth, Flash Next outperforms every other model that is available to us on the GB10 in all of my testing; though if you have two sparks then the TP=2 version is even better and easier (I don't think you'll need the nightly for that at all, just use the recipe)
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)
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.
Funny that all but one video has audio, the house 3D model one, where you can hear (what I assume are) Xiaomi's engineers talking about who knows what.
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.
Can we afford to look past it? If/when claudeslop starts infecting every new model to such an extent, that model will produce its own slop, infecting new models... At what point do we lose all reliable methods for establishing "truth"? This is epistemic collapse waiting to happen. I honestly thought it would take longer... holding out for a coherent shared reality in 2030 seems optimistic.
This is a big week. Probably getting next OpenAI and Anthro models, Grok 4.7, Mimo, etc. These open source model releases are why I can't take the "slow down" crowd seriously. I pitted older Mimo, qwen, step, gpt-oss, and other models against each other playing games like Werewolf and Sketch.io-like games where I let them talk shit while they played against each other. Mimo was by far pareto frontier of game-playing for the models that were <$0.15/m input tokens on OpenRouter. Qwen was pareto frontier in the shit talking game though. Qwen's hilarious. https://www.tiktok.com/@clankerfights/video/7642862917582425...
Leaning into what it cost to train is hilarious and an obvious shot at US frontier labs spending tens to hundreds of millions or more to train their models.
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
MiMo-V2.6-Flash-310B-A15B roughly GPT-5.6 Luna / Claude 4.9 according to benchmarks
MiMo-V2.6-Pro-1.02T-A42B roughly GPT-5.6 Sol / Opus 5 according to benchmarks
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.
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.
There are two (or more) concepts named after the same person:
- Pareto efficiency/Pareto curves: Basically the convex hull of points along the edge of a graph, indicating the best tradeoff between the axes. This is what the post is talking about.
- Pareto principle: this is the 80/20 rule you're talking about
This is the Pareto Front [1], rather than the Pareto principle. It's the idea that anything that's more intelligent is more expensive and anything that's less expensive is less intelligent.
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!
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.
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.
* 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.
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
Chinese companies are well documented in their copying / following / training on output / outright stealing tech.
China, as far as I know, has not come up with anything on their own and has only been able to play catch up and the memory optimization game as far as innovation.
OK, so how does this help the US?
If the US slows down this may lead to people that would have went to US labs to go to other countries.
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 people with enough money are more "persuasive"
I don't think "putting an upper bound" was OPs phrasing?
That's what pacing the frontier is, and is what the labs are pushing for.
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.
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.
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 ...
Please do
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.
the existence, who else has a live dashboard for the RL late-training?
> 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).
Pelicans for Flash: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
Pelicans for Pro: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
I think we can say pretty confidently they aren't pelican-bench-maxxing
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.
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.
Apologies for not taking the time to find it, but there was a post that tried to determine if the pelican was benchmaxxed across a bunch of models by comparing it to other SVGs, and found that it wasn't at all.
How does this translate to coding performance, which is what most of HN cares about (...I assume)?
It means they're good at writing SVGs, in particular SVGs of animals riding modes of transport!
I only visit HN for the pelicans, personally.
Yeah, I thought the "N" was for Nest
ish… at least we can be sure they don’t benchmaxx the pelicans lol
Anyone else more excited about Chinese models than American models these days? Big thing for me is affordability.
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?
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.
Not sure about that.
Given the difference in compute, it seems plausible.
However, the researchers at the US labs are surely no less talented, and they have better access to hire talent globally.
They too have to serve their models efficiently at a large scale, and with current capacity constraints this must be a top priority.
> 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.
> 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.
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.
API rates still aren’t quite competitive with the OpenAI x20 accounts, but they are definitely getting close with deepseek 4.1 flash. I spent a few days with only 4.1 and was very impressed.
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
Yes, an expensive American LLM has zero capabilities as far as I’m concerned because I’m never going to pay for it.
No, because I'd rather not support our economic and military rivals.
I'm Canadian so this sentiment has little value in 2026 unfortunately.
Also, frankly, as a fellow Canadian it's pretty clear that the biggest "rival" the US has right now is itself. Just passed out in the corner puking on itself shouting about all the foreigners who won't talk to it.
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.
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.
Canadians warming up to China makes me think of Germany becoming increasingly reliant on Russia in the 2010s.
Murica just has a MAGA problem. We can still be friends if and when you sort that out. Us Canadians like most of you quite a lot.
Yes, someone can still blow up a pipe and they look the other way. On the other hand, you can also draw parallels to themselves becoming increasingly reliant on US vs UK in the past.
Does it count as supporting a rival if your an American using an American inference provider self-hosting an open weight model from a Chinese lab?
Agreed, and also because I support freedom of speech!
Neither the US nor the Chinese companies are on your side then. They both censor, just different topics.
But at least I can run Chinese models locally, and strip a lot of that censorship/refusal.
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
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?
It is a 9B agentic model developed by Xiaomi MiMo through supervised fine-tuning of Qwen3.5-9B on MiMo-generated data
curious why the HF pill (on the right) always has inaccurate values
I noticed the same, and I wonder as well.
I suspect they are calculating something in the weights or config, I see it pretty consistently with quants
more like 500B in FP8
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).
My guess is that by function they break down frontend sections or components into pieces and I believe document things for themselves on some level, or purposely are verbose in this way. It is probably also shaped by users and existing web patterns. They probably get reinforced by models the more common they become.
Nice catch!
The extraneous small-caps labels are one of the main idiosyncrasies of AI generated markup. I wonder how much of this is a "scaffolding" technique to help the model build stable designs. But was it reinforced in RLHF or an emergent behavior of the models?
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.
wow im sure you were the only one saying this
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 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
ExploitGym DeepSWE v1.1Why not? In my own benchmark Opus 5 does in fact come out on top[1]
1 - https://bench.killswitch-lang.org/
Maybe you should not trust any of the benchmarks!
They match my experience. Astra and Fable I rate below Sonnet. They are incredibly poor. They were excellent for a couple of days after release and then plummeted.
Maybe I am being routed to more quantised versions or less capable models with system prompt to fake Astra or Fable.
I've got a working recipe to run this model on Dual DGX Spark: https://github.com/volfco/spark-vllm-docker/blob/main/recipe...
Averages ~25-35tok/s which isn't bad for a first attempt.
Conspicuous that there’s no reference to GLM 5.3/Flash in the reported benchmarks. Just Deepseek and Kimi.
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…
https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B is an option
That’s for toy GPUs, like the 5090.
there are many tasks (increasingly more each day) where small models are more than enough
Is there a gamechanger around the corner to reduce DRAM requirements?
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".
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
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)
https://github.com/spark-arena/eugr-recipes/blob/main/recipe...
This one!
I'd recommend pointing your agent at it (after installing sparkrun), and asking it to research the absolute latest in TP=1 Flash-Next - mine grabbed particular vLLM nightlies and mods to improve performance, and it was well worth it.
I have a quirky vLLM on k8s on 2x OEM sparks setup with 9 models available to me. I'm not keen to run nightly vLLM, too many issues with it in the past. Going the qwen-next path means displacing things I use daily :/
I have a watchful eye on the diffusion ~ Jev/Kev PR
https://github.com/vllm-project/vllm/pull/57250
For what it's worth, Flash Next outperforms every other model that is available to us on the GB10 in all of my testing; though if you have two sparks then the TP=2 version is even better and easier (I don't think you'll need the nightly for that at all, just use the recipe)
I'm so tempted to buy a second one...
prices have gone up quite a bit...
I'm running embedding, reranking, and policy tuned models too, and a Jev/Kev when that's landed. Flash Next is not a substitute for those
I have OpenCode/Fireworks to access big models
looks like this is likely it
https://github.com/spark-arena/eugr-recipes
https://github.com/eugr/spark-vllm-docker
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.
I personally stopped caring as much about the tok/s as the agents are largely in the background, and so have also moved preference from MoE to dense
I want to see about fine-tuning these models a bit on the GB10 to tame that over thinking and some other behaviors (like using tools I don't use)
qwen 3.8 seems to have been trained with some `rkt` that messes with tool outputs to "save tokens"
check out the spark arena website, its the raison d'etre
interesting, peer comment seems to indicate this is a possibility as well, will have to take a deeper look
n-grams can be kept on SSD, no need to hold them in any kind of RAM (at least w/o batching)
Am I missing a joke? WTF is URAM?
unified memory, not sure if anyone uses URAM, I human hallucinated it
>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
Funny that all but one video has audio, the house 3D model one, where you can hear (what I assume are) Xiaomi's engineers talking about who knows what.
This looks great in terms of cost and capabilities, truly pushing the frontier forward in terms of open weight light weight models.
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.
Can we afford to look past it? If/when claudeslop starts infecting every new model to such an extent, that model will produce its own slop, infecting new models... At what point do we lose all reliable methods for establishing "truth"? This is epistemic collapse waiting to happen. I honestly thought it would take longer... holding out for a coherent shared reality in 2030 seems optimistic.
how do you recognize "claudeslop"?
It's an honest, load-bearing, simple thing.-
It tops the intelligence vs time Pareto frontier and, uniquely for a Chinese model, does well in the E2E response time too.
https://artificialanalysis.ai/models/mimo-v2-6-pro#intellige...
This is a big week. Probably getting next OpenAI and Anthro models, Grok 4.7, Mimo, etc. These open source model releases are why I can't take the "slow down" crowd seriously. I pitted older Mimo, qwen, step, gpt-oss, and other models against each other playing games like Werewolf and Sketch.io-like games where I let them talk shit while they played against each other. Mimo was by far pareto frontier of game-playing for the models that were <$0.15/m input tokens on OpenRouter. Qwen was pareto frontier in the shit talking game though. Qwen's hilarious. https://www.tiktok.com/@clankerfights/video/7642862917582425...
Finally a lab that doesn't cheat on the charts
They mixed up DeepSeek 4.1 Flash with something else on this page, possibly DeepSeek 4.1 Flash means Gemini 3.8 Flash.
These benchmark are useless as they don't say whether they were done before or after Fable and Astra got nerfed.
Leaning into what it cost to train is hilarious and an obvious shot at US frontier labs spending tens to hundreds of millions or more to train their models.
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
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
Pretty sure that's Luna xhigh.
As for the stats that everyone wants:
MiMo-V2.6-Flash-310B-A15B roughly GPT-5.6 Luna / Claude 4.9 according to benchmarks MiMo-V2.6-Pro-1.02T-A42B roughly GPT-5.6 Sol / Opus 5 according to benchmarks
Perhaps with IQ2 flash will run on 128G M5?
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?
yes. I use opencode and opencode uses Xiaomi as a provider.
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.
There are two (or more) concepts named after the same person:
- Pareto efficiency/Pareto curves: Basically the convex hull of points along the edge of a graph, indicating the best tradeoff between the axes. This is what the post is talking about.
- Pareto principle: this is the 80/20 rule you're talking about
No, Pareto refers to Pareto efficiency https://en.wikipedia.org/wiki/Pareto_efficiency
What you call "frontier line" is also called "Pareto frontier" https://en.wikipedia.org/wiki/Pareto_front
Your description of it is basically correct though
This is the Pareto Front [1], rather than the Pareto principle. It's the idea that anything that's more intelligent is more expensive and anything that's less expensive is less intelligent.
[1]: https://en.wikipedia.org/wiki/Pareto_front
"Pareto" is many things, but here it does indeed refer to the frontier: https://en.wikipedia.org/wiki/Pareto_front
Thank you guys. I learned something new.