MiMo-v2.5-Pro-UltraSpeed: 1T model with 1000 tokens per second(mimo.xiaomi.com) |
MiMo-v2.5-Pro-UltraSpeed: 1T model with 1000 tokens per second(mimo.xiaomi.com) |
Really?
I think this site often overlooks that second group and how large it likely is.
It's such a weird "Gotcha" that seems to only assume that Chinese LLMs might censor something.
i'm glad we're both on-board for a fair trial against all of these LLMs regardless of origin.
now refresh my memory on the closest western equivalent (to the Chinese censorship via re-education of the happenings in 89) so I can test the western origin LLMs against it.
Say, I work for Planned Parenthood and want to use a LLM to help me develop code. Will it refuse to run because there are mentions of abortion? Everyone has a different censorship line, but unfiltered is more generically useful.
Anything different for Grok?
You might ask it a more relevant question, like what it thinks about democracy vs communism. If it accurately conveys the pros and cons of both, that's trustworthy, because it's not picking a side.
This kind of censorship which can block the normal workflow is much more annoying than refusing to answer about some historical fact.
Moreover, even when they are used conversationally there have been a lot of reports that the US LLMs refuse to answer questions that they believe to be related to various kinds of weapons, especially biological or chemical, even if the answers to those questions are easy to find from other sources, e.g. from Wikipedia.
Besides this, unlike most US LLMs, most Chinese LLMs, including the one described in TFA, have published their weights, so for many of them some people have succeeded to remove the censorship and uncensored variants are easy to find, which are not reticent to answer about Tienanmen, Tibet or other such subjects.
What actually matters is that the mere tool is withholding information at all, and that the boundaries were set by whoever designed it.
Dont get me wrong I've been an advocate of this stuff (I carry two phones, one with GOS for my personal use and the other for ID verifications). However, without reasoning, you just can't see it, because you're as biased and propagandized as anyone in China.
It's also pretty funny sometimes how it gives weird future roadmap estimates ("part 2 - 3 weeks, part 3 - 2 months", etc.) and when you tell it to actually do those changes it's pretty much done in half an hour
https://openrouter.ai/deepseek/deepseek-v4-pro?sort=throughp...
(I should go measure this now, I'm curious)
It will go much faster.
There can't be many normal use cases where there'd be any cost benefit.
It's a cute toy right now, but you can tell an LLM that it's an http server, and have it respond directly to a web browser hitting it. It generates headers in response, as well as page contents. As 1000 tok/sec becomes three new normal, we will come up with newer ways to use it outside of toy fiction encyclopedias.
So long as AI lives in server farms, humans will be needed for tasks in the physical world.
It's only if we combine AI with robots that things get really dicey.
So, if any, I would say it's worse for us. Obviously, it's the completely opposite situation for corporations and executives: they are loving the AI situation so much!
Data at https://gertlabs.com/rankings
It is another thing the the BigLabs accuse open weight models of benefitting from distillation & other techniques & essentially avoid higher training costs (which typically bleed into bills end users pay for inference).
Ex A: https://www.anthropic.com/research/2028-ai-leadership
Ex B: https://www.reuters.com/world/china/openai-accuses-deepseek-...
Discussions about choosing a library with the best syntactic sugar method naming is just as crazy as suggesting we type in assembly.
For a while I was running Cerebras GLM 4.7 for a bunch of tasks. Not a very smart model, but it's fantastic to be have a live prototype of a site up and be able to type "make the fonts bigger. No not that big" and see it change in real time. And MiMo 2.5 is a lot more capable than GLM 4.7.
MiMo 2.5 is not the same model as MiMo 2.5 Pro.
GLM 5.1 is z.ai's lastest iteration & is one of the popular open weight coding models.
If you've had the chance, how does GLM 5.1 (which is now more expensive than MiMo 2.5 Pro after its recent 70% price drop) compare?
But quite a bit more expensive than MiMo 2.5 Pro. Like 5x to 10x more on my little tests, at least by the API rates.
The Xiaomi team really brought something to the table.
I think the answer is that there's a tradeoff here where additional throughput for a single person can be achieved only by tying up more resources than a normal request would, even when you take into account the fact that the normal request takes longer to finish. I'm not an expert, but some of the optimizations they describe, particularly the parallel prediction stuff, sound like they could take up extra resources.
I think the margins are getting quite compressed with this one, since it isn't included in token plan and the actual costs increase are much higher than just 3x. But still fairly decent.
Remember, these guys are not VC backed. Anything they do must break even
Understand the spirit of this, but probably not true. I don't think Xiaomi, or any big tech company, needs to break even on their new model releases.
From that point of view, they have as much money as they need. That's why there is no "VC", because Chinese government assumes that role.
Despite the performative UI components they have a shipped (demo) product:
This is only 3.1 8B and a very small context window, but at 17k tokens per second it's likely enough to reliably call tools which would make a huge difference in agentic applications. Assuming they can bake in better models I'm just as bullish or even moreso on this, considering this opens up edge computing at the extremely low power requirement.
High tok/s is the future IMO.
edit: now I read the article fully, seems like they utilize some very effective MTP algorithm. and somehow the quality is still decent enough.
though, I doubt that the quality really only drip a bit like they claimed. maybe for the benchmarks, but for general uses the heavily quantized models very often so worse result.
- persistent CUDA kernel
- tiled processing with overlapping read/writes
- model designed with specific constraints in mind
Getting ~1000 TPS on near-frontier intelligence is a step change, and enables whole new use-cases for applications. Seeing limited compute resources beget selective access makes me worry for the future of competition.
Are you kidding me. Come back when you are ready for the users. I was hopping to try it, what a frustration.
128 sounds really tiny, I wonder if they mean some kind of blocks?
[0] https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro-FP4-DFlash#4...
I’m excited for ultrafast AI. It likely means less temptation to multi-thread and deeper flow in single sessions.
First make it write a contract (REQ/ARCH/IMPL documents). Skim through those for any mistakes.
Then based on those ask it to write tests. Again skim through them.
Now you have a context full of guardrails. It’s less likely to surprise you.
If you're treating it like a slot machine you're doing it wrong. It will give you exactly what you ask for if you ask clearly, i.e. write a clear, detailed specification, not just "do X!". The nondeterminism comes from vagueness in specification.
i've a Github copilot yearly subscription. Microsoft recently changed their billing to based on token. i'm still getting billed per premium request but GPT 5.4 is now 6x compare to 1x before.
I genuinely don't understand what moat these US model labs have. If they're saying recursive self improvement is just around the corner and Chinese labs are only slightly behind the leading US models, what moat does the US labs have? Are the US models going to recursively self improve better than the Chinese open source ones or something?
I might be completely wrong about this, but if I had money in OpenAI or Anthropic I'd be pulling it all right now. I think the chance of them going to near-zero over the next few years is very significant.
Or Google. I'm working with multiple customers right now that are very pissed at Google for deprecating Gemini 2.5 Flash, canning the GA release of 3.0 Flash and now have to decide whether to bite the bullet of the 5x price increase for 3.5 Flash or switching providers. Quite a few of them will likely fully pivot to open models.
For non subsidized plans? Pretty sure they'd need to put this in ToS, or law suites would have followed by now.
This strategy will seem to work really well until the economy that enabled that foundation to form is hollowed out. Then, there will be a reckoning (but we will have no choice but to march forth from there).
Especially as teams invest in proper agentic harnessing.
We have had a champion in our team that has invested a lot of time into it over the last 4 months, and if anything, quality has improved, not decreased. Architecture is more coherent, codebase has been cleaned up, agents find information quickly, code produced is very solid and my role is more and more checking that the output meets the requirements. But I cannot confidently say that I would've done a better job than AI more often than not I have to admit it does a better job than mine.
The mistakes are less and less technical and merely in the domain mapping. And AI is still not creative as I am for finding solutions quickly to unlock stakeholders' issues. Also, AI is still not creative as I am for finding the proper solutions for advanced technical problems. But it does a better job than me, even on that front, one shotting few solutions in a fraction of a time it would've taken me to test one idea myself.
Mind you, I don't like AI and I think it ruined the job, I don't like working this way, it's exhausting, way more work on one side, way less fun and fiddling with technical parts.
And yet, I have the genuine belief that few years from now we'll be cloning open source repositories that are already optimized/harnessed and tested for agentic loops and best practices left and right with software engineers mostly overseeing the domain translation and putting their 2 cents on the non-boilerplatey parts of the product (which, in general, are a small part of the surface).
I think that the next years of my career will be mostly spent in setting up and writing the harnessing and domain mapping part. Then I will move to another sector, not because I necessarily believe I won't have a job, but because I want to vomit thinking that's going to be my job.
> No one cares anymore.
I never cared about this.
I think this captures something that I've been searching for the words for. (Maybe I should have gotten an LLM to write the words for me.) Some of the biggest AI boosters are the kind of dev that would have cared about the new frameworks of the last 3 months. They had a "the framework does all the thinking for me" attitude already, so it is easy for AI to slot into that.
it needs to win marketing landscape, hyper-overcrowded by thousands of competitors, slop-gened over weekend.
I have a more hopeful take. As AIs improve and get faster we can more quickly and iteratively improve code which we may have historically avoided due to the work involved.
I know i've made several refactors that would have otherwise been insane lifts. Not only because the work involved but because sometimes you don't know if it will work, and so you have a sort of double friction; you don't know if it will even succeed. With an AI you can just throw it at the refactor to see if it runs into a problem all while you're having a coffee break or w/e.
In general AI is going to enable humanity to be more extreme versions of itself. For good and bad. I suspect more bad than good, though.
> It uses 384 routed experts (top-8) with hybrid attention (full-attention + sliding-window 128 at 6:1 ratio) over 70 layers (1 dense + 69 MoE)
I'm not saying there aren't any use cases for super-fast (and super-expensive) generation, but it does seem a bit niche. If it was free then sure faster is better, but what are the mainstream use cases where people might pay 3x more for a faster version of something that is already fast?
I think it would have to be an application where it paid for itself - where the 10x faster response was actually worth more than 3x the cost to you - where the extra speed was worth the extra cost.
Sometimes Opus just gives me a rubbish session.
In software + GenAI now every housewife can build some App over evening.
"Watching John with the machine, it was suddenly so clear. The terminator would never stop. It would never leave him, and it would never hurt him, never shout at him, or get drunk and hit him, or say it was too busy to spend time with him. It would always be there. And it would die to protect him. Of all the would-be fathers who came and went over the years, this thing, this machine, was the only one who measured up. In an insane world, it was the sanest choice."
As long as you've indicated what you want, the machine will try to do what you ask of it. It won't get tired because "the codebase is too big", or it has gotten bored of the pattern, or it wants to introduce a new technology.
It just does the thing you asked of it. (note, that yes, I get that as a codebase size increases, it might make it more difficult to fit into context, but that only applies if it needs to read a large percentage of the project to implement the task, which shouldn't be the case.
there are good actors, which are empowered by AI to produce positive impact, but often there are N times more bad actors, which push crappy code to close feature requests fast, increase performance LoC-like metrics, etc.
I'm not agreeing or disagreeing with you, but my brain cannot comprehend how machines can advance such interconnected systems while keeping humans in focus.
Perhaps I shouldn't have watched the Animatrix again.
There will only be a reckoning if models don't get much better.
If they do get much better you can just have them refactor, fix bugs in, or replace the existing codebase.
The concept of tech debt is sort of meaningless if you anticipate intelligence gains in models to continue.
In this case, at least it’s threatening multimillion dollar salary jobs instead of entire towns of working class people in America or Mexico.
And the Chinese labs actually release their weights. You could call it… open AI.