I been using MiMo-V2.5 to do most of my work as software engineer, on a variety of projects I'm working on, and I been VERY happy with ROI. The model is very powerful! Not perfect – I've run in hallucination loops once or twice, but nothing a stop-then-continue wouldn't solve.
The cost is unbelievably low, and the quality of intelligence I get is equivalent to when I was working mostly with Anthropic models (late last year/early this year). I'm fully invested in MiMo and I'm very happy with it.
--
PS: I also check almost daily to see if other models are capable of doing such great work. And they do – DS4F is powerful and DS41 is impressive, GLM 5.3 Flash gets a job done well, etc. – but when I add cost of M-token in the ROI math, Jeez! MiMo is an order of magnitude better.
That's an eternity when it comes to coding models.
In my personal experience, we've had almost a step change every ~3 months this year, at least for bigger one-shot tasks. For example looking at Gemini Flash 3.0 vs 3.5 vs 3.8, it went 5% -> 30% -> 75% on DeepSWE, all since the start of the year.
In another one, Opus 4.6 level already solved 90% of my work-day tasks, so while better models have been instrumental into handling a higher % that does not mean that defaulting on cheaper models can't be good.
I run DS 4.1 flash daily, and then cross check with gpt-6-astra and I've nuked 90% of my AI monthly bill while having higher limits and better performance/intelligence than I did just at the beginning of this summer.
Because they’ve essentially exhausted pre training scaling and are looking to post training to expand capabilities, which is really just optimization via reinforcement learning against specific tasks aka bench maxing.
Same here. It’s the first AI provider I actually gave money to, since they offered the model for free with a Mimo code for the first month or so, and it was great.
These days, there are more intelligent models like DS4.1, but Mimo is very obedient, so I plan things with another model and give the implementation to Mimo.
I've found that mimo v2.5 works for very basic things like a python script to do one thing, but it also is very 'dumb' compared to qwen 3.8-flash-next (I think the benchmark scores for terminal and coding specific benches back this up). And definitely not in the same class as like a GLM5.2 or 5.3. It's fast but makes basic mistakes that only get caught later.
The fact I can run Qwen 3.8 Flash Next locally, forever (on my DGX Spark-alike) is genuinely shocking to me. It’s crazy good for how small it is. Fast, too.
Yeah, I'm guessing you have a variant that fits in <128GB with 262k context? I have the unsloth Q8 GGUF of it here in a setup that with full context and ton of extra llama-server "--cache-ram" sits around 200GB RAM usage on a 256GB system, it's probably the best thing I've found for a 256GB class machine. Enough headroom for a rope/yarn extension to 524288 context if I need it.
RTX6000 Blackwell with 96GB is enough to run it with NV4, 256k context, KVcache, multimodal at 130t/s (SGLang). It's toasty, you're using up 94GB of those 96, but it works and the results are great
For comparison, I am currently at 6.6B tokens, 95% of monthly quota on a 10$ command code plan, mostly using DeepSeek flash 4.1, or some of the free models for easier tasks.
You use the low cost Mimo-V2.5 and not its big brother Mimo-V2.5 Pro? I also made good experience with Mimi-V2.5 when used in conjunction with prewalk mode. But then other models got so cheap and perform better, so I only use Mimo for background tasks.
The availability of Mimo over Openrouter got however, much worse recently.
They both are in the 50-100 tok/s range. The Mimo v2.5 Pro Ultraspeed beta could reach 1000 tok/s, hoping they can do something similar for the new model, it was amazing.
mimo 2.5 has been a big underperformer since shortly after it's release imo. i cancelled my sub after the first month. purposefully using 2.5 right now is just handicapping yourself for no reason.
Yeah, this Brazilian dude who has been a contributor here on HN longer than your anonymous account is shilling for a Chinese model company. Makes sense.
For my own usage, Luna is cheap enough that I don't care if other models are cheaper. I'm interested if another model is in some way better and not too expensive.
Luna is great but makes a lot of mistakes at high and lower in my experience (large rust codebase). I use Luna Max for asynchronous subagent reviews and am very happy with its work, but it’s slow af.
DeepSWE is saturated now IMO, and is basically worthless. Lots of new models get around 74%. Shame too, because it was a pretty decent benchmark for a few months there.
Also worth taking a look at is the mimo harness. It's a fork of opencode with some new modes added for long horizon tasks. One of the better open harnesses out there at the moment.
> and google just started letting all their engineers use claude
That's misleading.
1. Having different models available is useful for A/B testing and helping improve Gemini itself.
2. They have an enterprise offering for Antigravity (their agentic coding platform), and they need to test that it works well with non-Gemini models too.
Neat! I've been trying out their next model for the last week, which I assume is a version of this, and it's been a good experience so far.
I had used 2.5-pro for a hefty chunk of development, and found it to work like a somewhat forgetful senior engineer who was new to my project. Very capable, would almost always choose a reasonable option, if not always the best one for the project, and not great at multi-tasking. Generally, made me comfortable not scrutinizing the code line-by-line, but still needed a bit of steering once projects got to a reasonable size.
The next model is a clear step up in the multi-tasking capability at least, with me very rarely having to steer the implementation of a well-defined issue. In terms of code, I found MiMo-V.2.5-pro to be extremely conservative, implementing minimal solutions. The next model seems a little bit more ambitious, in positive ways, making good guesses about gaps/next steps. It also seems to be a fair bit better at design, at least for the little bit I've done, it was good at translating my concepts to practical elements on screen, and cleaned things up nicely as I made suggestions.
Speculating here, but I assume researchers can make a reasonable estimate of the size of closed models based on factors like training time, training speed, and the number of tokens processed.
Also, Anthropic and OpenAI probably want to keep each other on their toes so they don’t end up on the wrong side of another Opus 4.6 / GPT-5.3-Codex situation, where one lab releases a model only for the other to drop a better one hours later.
I think first of all it’s not an obvious idea, also the marketing surplus for other providers is not as big for openai/anthropic as for xiaomi and last but not least I’m pretty sure you can withdraw methodology from here.
I’m saying who has a million dollars for me, so I can make my own model?
Kinda yes. The benchmarks become part of the validation set, which means the models get slightly overfit to them if they are used as criteria for stopping the training. But a lot less compared to using them in the training data.
I'd guess everybody uses at least some benchmarks as stopping criteria, which is kinda sensible, but it also does induce some benchmaxxing, and explains partly why the newest models always tend to eke out in benchmarks.
Correct. If just stopping criteria, that is less contaminated. The question gets muddier once you also use it to determine hyperparameters during small-scale runs.
They are using it to evaluate checkpoints during the training, they are probably not using the benchmarks for training the models. It's a common practice for big reinforcement learning runs.
They exist to detect degradation. Datasets are not perfect and if a batch contains too much bad data it can ruin a run, also an opportunity to find bad data and improve the dataset filtering.
They run one step/iteration on an additional chunk of training data, then use the snapshot of the weights after that iteration in a separate validation benchmark while continuing to train on another chunk of data for the next iteration.
They result of the benchmark does not feed back into the training, it simply serves to provide a measurement of progression over time.
You gotta have something to aim at. And, presumably, the benchmark is not part of the training data, it is the test against which the model is tested at each stage; is behavior moving in the right direction?
It’s the difference between “study law until you can pass any random bar exam” and “here are 200 legal questions and we’ll drill them, with me correcting and explaining when you get one wrong, until you can pass exactly these 200”.
Your right that tuning can aim for a benchmark, but it does not leak any information about the answers.
The first implies generalization. It's not a test of generalization.
It's actually close to the second. "Here are 200 software questions, will drill you on *other stuff* until you can pass exactly these 200. If the other stuff isn't improving your scores we will change ratios of it till it does."
The reason it benchmaxes is that *other stuff* ends up looking more and more like SWE Bench without you realizing it.
This is crazy, but sadly anthropic/openai will never do this, what has happened to this world, where chinese companies are more open than US or even EU companies
I don’t think Trump changed them, but Trump is absolutely a symptom of larger social collapse in the US, and that collapse has affected Altman and Amodei. We’re not even pretending that truth matters or that the wealthy can ever suffer consequences, and those two seem quite liberated by that.
"As we get closer to building AI, it will make sense to start being less open. The Open in OpenAI means that everyone should benefit from the fruits of AI after its built, but it's totally OK to not share the science (even though sharing everything is definitely the right strategy in the short and possibly medium term for recruitment purposes)."
-Ilya Sutskever (email to Elon musk and Sam Altman, 2016)
I remember Zuck bragging about using a 100MW DC for training, and Musk's Colosus was supposed to be a training data center (but they fucked up the design so they had to repurpose it to an inference one).
Total run cost is $1.2M until now, what resources are they using to train their model? Wish they shared more details on that and what the MFU metrics are.
I absolutely love that someone is doing this! Why isn’t IBM for Granite or Google for Gemini?
If you are going to develop a near frontier model, and you don’t think you have special sauce up your sleeve, why not making training runs and RL environment scores etc. visible to the world?
I’m genuinely learning quite a bit just from the dashboard
I don’t see how it would head off such accusations. This is post-training, and even it’s data could be pulled from other models or run against other models in realtime. Not saying that’s the case, just that the dashboard does not disprove.
"Stealing millennium problems" would be more complete if not accurate description. And that 99.99% of the market is not interested in solving millennium problems is the other fact.
Haha yeah pretty wild how easily you can see the data is fake by the repeating numbers (refresh the page the progress goes back in time constantly) + watch for restarts. They say they happen but 0 data correlates the log messages. Just a replay of old data or being fed by an llm so they convince people they are open
Does it even matter if it distills other models? ethically no, because the information was originally stolen from us users... it is indirectly harming anthropic if it is happening, but who cares? It is better for everyone except anthropic that anthropic doesn't have a monopoly.
With other software, devs convince their managers of the importance of using open source stuff in their stack. With AI, it's usually managers choosing what models to use for the devs. The US labs don't need to give a damn how much devs like open source
This isn't about liking open source. This is about the labs just being cool and doing cool shit instead of the opposite which is Anthropic where all they talking about is killing everyone and taking everyone's job.
Well maybe they should because companies that do cool shit tend to attract people wanting to do cool shit and those are the people you want to be working at your company.
Unless you want your company to become Google which use to do cool shit but now they dont and now they don't have a single person at the company that can build cool shit so they just release a bunch of lame shit.
I was dev, and now I am Senior level manager.
Open Weight models are current main focus for many companies with full alignment with top management for very simple reasons:
- stable and predictable performance (no pre-launch models degradation)
- ability to tune them for specific business cases (though still rare tbh)
- better (at least 60% Opus vs Kimi (real,3rd party)) and more competitive pricing
- flat pricing if tokenusage is big enough to justify renting GPU
- decent quality
- much higher guarantees that data will not be sent somewhere (assuming 3rd party inference providers)
- and cherry on top: flat and minimal pricing with absolute confidentiality using Alibaba Apsara stack of recently released AMD Instinct Coder box[1]
> The US labs don't need to give a damn how much devs like open source
In the short term, true.
In the long term, unknown but typically when you hold progress that way while other countries don't you at best end up becoming siloed while the rest of the world continues on without you.
You mean all of the frontier models that the Chinese distillation clones are copying? Yeah kinda cool imo. If a dashboard showing training for a model that doesn't even come close to anything us labs have released in 6 months is "cool", then you're a loser
Anthropic and OpenAI literally stole from every human in history and youre out here complaining that the Chinese are distilling models and releasing them to the public?
Because without those labs to distill from the pathetic Chinese labs wouldn't have anything. Im not impressed by them copying US labs not sure why you are. But go off ccp bot
A restart of the process does not necessarily mean reverting the model state. I don't know why you would even do that, because you'd lose all the progress you made.
If you're thinking of the UI style, definitely not Claude. It is incapable of writing a clear sentence like "what each step's samples are made of", would have used all-caps for everything, more padding and gradients.
I hope this is /s because it’s very easy to get Claude to write sensibly. That’s why AI slop writing is so annoying because it’s so easy to avoid with any amount of effort at all.
In my experience Opus and Sonnet 5 subtly ignore most instructions related to writing style, and continue to sound the same half of the time. Do you have a successful skill/prompt to share?
My use case is generally easy to read instructions for lay people of an international/ESL audience. Have it write it's whatever and then run that on it and it comes out... actually pretty good. Use it for emails, etc, when it doesn't need a personal touch and just needs to be clear.
These skills won't get you a snazzy blog post, but I imagine could be augmented to produce something significantly better than the incomprehensible non-sense that it spews out by default.
It was about ten minutes of prompting. Honestly - oooh can we use honestly anymore? Not anything fancy - make a doc with ai tells, make a readability scoring script with popular measurements and weite a skill to use it skill, bring it all together with asd-ste-100. I don’t think I even did much tweaking beyond that.
They'd be running in the red then cause they charge way less than Claude. Sorry but it just doesn't make logical sense. They have open source, papers, and self hosting too
The cost is unbelievably low, and the quality of intelligence I get is equivalent to when I was working mostly with Anthropic models (late last year/early this year). I'm fully invested in MiMo and I'm very happy with it.
-- PS: I also check almost daily to see if other models are capable of doing such great work. And they do – DS4F is powerful and DS41 is impressive, GLM 5.3 Flash gets a job done well, etc. – but when I add cost of M-token in the ROI math, Jeez! MiMo is an order of magnitude better.
reply