Mistral Medium 3.5

https://mistral.ai/news/vibe-remote-agents-mistral-medium-3-5

by meetpateltech • 3 months ago

500 216 3 months ago

mistral.ai

amunozo7• 3 months ago

I want to believe it's gonna be good, but after trying GPT-5.5 even the most advanced Chinese models seem depressing.

ako• 3 months ago

Then you’ll be happy to learn it’s not Chinese

dotancohen• 3 months ago

GP is stating that the second best in the field, the Chinese, is so far behind the best in the field, GPT 5.5, that it is not even worth testing anything else.

r0b052• 3 months ago

This is a French model sir

spwa43• 3 months ago

Évidemment

Funny detail: Google AI (the one they use in search) can't spell évidemment correctly.

lava_pidgeon• 3 months ago

Honestly I depends on the context which this performance matters. Mistral is quiet cheap

manishsharan• 3 months ago

I am not following this obsession with SOTA and benchmark rankings

I have been using DeepSeek and GLMnmodels with OpenCode and Codex and Claudr side by side.

I have not found the Chinese models lacking. I enjoy for coding and like to maintain full control of my codebade and deeply care about the GOF patterns. So I am very stringent in terms of what I want the LLM to code and how to code.

So from my perspective, they are all about the same.

amunozo7• 3 months ago

That I agree with, but for more complex autonomous changes the differences are considerable. However, it seems that most models will reach the saturation time in which they will be useful for almost everything and the difference will be in more and more niche and specialized tasks.

InputName• 3 months ago

Looks at first graph. It's SWE-Bench Verified. A benchmark Open-AI stopped using two months ago due to contamination.

Doesn't look to promising. Is there any reason to consider Mistral other than it's not US?

tpurves• 3 months ago

If it's not US and it's within a few percent of SOTA that might be good enough for a lot of people (eg Europeans)

NitpickLawyer• 3 months ago

Gemma has been better for us at EU languages than mistral (for comparable sized models) :/ so ... dunno. What mistral does well and others are lagging behind is deploying on prem with their engineers and know-how, offering tuned models for your tasks and finetuning on your own data. (I expect google to start offering this next)

amunozo7• 3 months ago

Price and speed.

2ndorderthought8• 3 months ago

They did not stop using it due to contamination. They said it's flawed and indirectly said anthropics results were impossible. It's very possible they are sore losers

spwa43• 3 months ago

TLDR: Mistral Medium 3.5, text-only, 128B dense model, 256k context window, modified MIT license. Model is ~140G ...

https://huggingface.co/mistralai/Mistral-Medium-3.5-128B

They more or less claim this exceeds Claude Sonnet 3.5 on most things, but is worse than Sonnet 3.6, and exceeds all other open models.

Oh and they have a cloud service that will code your apps "in the cloud". But, yeah, at this point, so does my cat.

And, yes, unsloth is on it: https://huggingface.co/unsloth/Mistral-Medium-3.5-128B-GGUF (but 4bit quant is 75G)

Marciplan• 3 months ago

You mean Sonnet 4.5 and 4.6 riight

spwa43• 3 months ago

right

wolttam• 3 months ago

Sonnet 4.5 and 4.6*

There is no way it exceeds “all other” open models - but it does exceed all of Mistral’s past models.

You can see it getting blown past by GLM 5.1 and Kimi in this.

Still excited to give it a try

2ndorderthought8• 3 months ago

It looks like qwen 3.6 is winning and smaller for the April small model roll out

pama3• 3 months ago

Unfortunately they only compare to old “all other open models”. There are probably over 10 other open models better than it by now.

mtct88• 3 months ago

It's okay, nothing exceptional, but any news from non US and non Chinese models is still good news.

pb73• 3 months ago

This is the bar for Europe, huh?

amunozo7• 3 months ago

This is the bar for anybody that's not the frontier labs.

locknitpicker3• 3 months ago

This is the bar for Europe, huh?

A few months ago China was being criticized left and right on how somehow it was not able to compete, and once DeepSeek showed up then all the hatred shifted onto how China was actually competing but exploring unfair competitive advantages.

Funny how that works.

Also, aren't the likes of OpenAI burning through over $2 of investment for each $1 of revenue?

deaux7• 3 months ago

Where are the competitive models from Singapore, Japan, Taiwan, Korea, Russia, Canada, India, the UK? From anywhere that isn't China or the US?

There are none. Mistral Small 4 is pareto-competitive in its pricing bracket at $0.15/$0.60, at worst it's second to Gemma 4 26B A4B. The above countries have never had a model that is even close to being so.

This particular Mistral Medium looks to be uncompetitive at that pricing. I'm surprised it's so expensive given its size. Wonder if we'll see other providers offer it for cheaper.

but that doesn't mean Mistral has never produced anything useful.

Matl5• 3 months ago

I mean, at least we're not melting the planet trying to predict the next token that sounds about right.

wyre• 3 months ago

I'm rooting for Mistral. It seems they are making a big bet that smaller models will win over larger ones and I can see it happening. I was running some simple chat and tool-calling benchmarks for small models and Mistral Small 4 performed well for it's price ($.15/$.60). Seeing this today got me excited, benchmarks seems solid compared to models much larger, but it's priced higher than Haiku, 5.4 mini, and all the the Chinese models it's comparing itself too. It's not even winning those benches either, just being competitive with them, which is great, those models are 5x+ the size, but they are also 1/2 the price. Hard to be excited about that.

postalcoder4• 3 months ago

This release Mistral really reminds you of the gap between the frontier labs and everyone else.

Pre-agent, there wasn't always an obvious difference between models. Various models had their charms. Nowadays, I don't want to entertain anything less than the frontier models. The difference in capability is enormous and choosing anything less has a real cost in terms of productivity.

I've been a big fan of the smaller labs like Mistral and especially Cohere but it's been a while since I've been excited by a release by either company.

That said, I'm using mistral voxtral realtime daily – it's great.

onlyrealcuzzo• 3 months ago

Pre-agent, there wasn't always an obvious difference between models. Various models had their charms. Nowadays, I don't want to entertain anything less than the frontier models. The difference in capability is enormous and choosing anything less has a real cost in terms of productivity.

It's just apples to oranges.

There is not a clear, across the board, winner on non-agentic tasks between Gemini, ChatGPT, and Claude - the simple chatbot interface.

But Claude Code is substantially better than Codex which itself is notably better than Gemini-cli.

In this vein, it should not be surprising that Claude Code is way better than non-frontier models for agentic coding... It's substantially better than other frontier models at specialized agentic tasks.

nothinkjustai• 3 months ago

CC is not better than Codex, nor is it better than OpenCode, Crush, Pi etc…

postalcoder4• 3 months ago

I think there's a fair amount of evidence that the heavy harnesses actually drag down performance compared to bare harnesses.

philipbjorge• 3 months ago

I’ve been comparing Claude Code and Codex extensively side by side over the past couple of weeks with my favorite prompting framework superpowers…

From my perspective, Claude Code is decidedly not better than Codex. They’re slightly different and work better together. I would have no issues dropping CC entirely and using codex 100%.

If you’re working off of “defaults”, in other words no custom prompting, Claude Code does perform a lot better out of the box. I think this matters, but if you’re a professional software developer, I’d make the case that you should be owning your tools and moving beyond the baked in prompts.

locknitpicker3• 3 months ago

Pre-agent, there wasn't always an obvious difference between models. Various models had their charms. Nowadays, I don't want to entertain anything less than the frontier models.

This is a very naive and misguided opinion. In most tasks, including complex coding tasks, you can hardly tell the difference between a frontier model and something like GPT4.1. You need to really focus on areas such as context window, tool calling and specific aspects of reasoning steps to start noticing differences. To make matters worse, frontier models are taking a brute force approach to results which ends up making them far more expensive to run, both in terms of what shows up on your invoice and how much more you have to wait to get any resemblance of output.

And I won't even go into the topic or local models.

postalcoder4• 3 months ago

You need to really focus on areas such as context window, tool calling and specific aspects of reasoning steps to start noticing differences.

This is like saying "the current models and the old models are the same if you ignore every important advance they've made"

deaux7• 3 months ago

Can't agree at all. Productivity gap just 1 year ago was much larger for frontier model vs non-frontier. Let alone 2 years ago.

postalcoder4• 3 months ago

When I was thinking pre-agentic, I was actually thinking more pre-"coding seen as the main use case for these models".

2ndorderthought8• 3 months ago

Same. The gap is almost paper thin for anyone who hasn't gone full uninformed vibe code.

vessenes• 3 months ago

As always, rooting for these guys — model and national diversity is great. This looks like a solid foundation to build on; hopefully the 3.6/3.7 will dial in more gains. It looks like maybe from the computer use benchmarks that their vision pipeline could use improvement, but that’s just speculation.

The different results on some benchmarks vibes as if this is truly an independently trained model, not just exfiltrated frontier logs, which I think is also really important - having different weight architectures inside a particular model seems like a benefit on its own when viewed from a global systems architecture perspective.

Tepix• 3 months ago

I use Mistral Le Chat quite a bit.

One thing in particular I was disappointed in was its bad explanations when asking about French grammar. It made multiple mistakes and the other models got it right, even Qwen 3.6 27b!

Anyway, I'm hoping they catch up some more.

kubb• 3 months ago

There's a good chance that they'll catch up. The "AI race" is a race to the bottom, with the leaders blowing huge wads of cash on capabilities that get replicated months later by the competition at a fraction of the cost.

The only benefit of leading is mindshare. OpenAI is doubling down on that, by investing in communication companies. That's their pathetic attempt at a "moat".

pb73• 3 months ago

They catch up by distilling frontier models. They will eventually figure out how to prevent that from happening. No one has any interest in investing tens of billions if the product can be copied and sold for less.

mark_l_watson2• 3 months ago

I like the idea of Mistral, but the last time I evaluated Mistral Vibe it was really nice for $15/month but not as effective as Gemini Plus with AntiGravity and gemini-cli. I am currently running Gemini Ultra on a 3 month 'special deal' and AntiGravity with Opus 4.7 tokens is pretty much fantastic.

That said, when I stop spending money on Gemini Ultra, I will give Mistral Vibe another 1-month test.

I like the entire business model and vibe of Mistral so much more than OpenAI/Anthropic/Google but I also have stuff to get done. I am curious if Mistral Vibe for $15/month is a stable business model (i.e., can they make a profit).

amunozo7• 3 months ago

I'm testing it right now and it seems very buggy and unstable, just like before.

danelski2• 3 months ago

How do you feel about the responsiveness of gemini-cli? I tried it on a paid plan and the 10-minute hang-ups (per step, not the whole plan execution) really break the illusion of performance gains, unless you run it in the background and do something else in the meantime. It's more noticeable when Americans are awake.

mark_l_watson2• 3 months ago

it is usually fast, but if gemini-cli or any other coding agent is sluggish I quit using it for a while.

simonw2• 3 months ago

I can't figure out if this is available in the official Mistral API or not.

Their model listing API returns this:

  {
    "id": "mistral-medium-2508",
    "object": "model",
    "created": 1777479384,
    "owned_by": "mistralai",
    "capabilities": {
      "completion_chat": true,
      "function_calling": true,
      "reasoning": false,
      "completion_fim": false,
      "fine_tuning": true,
      "vision": true,
      "ocr": false,
      "classification": false,
      "moderation": false,
      "audio": false,
      "audio_transcription": false,
      "audio_transcription_realtime": false,
      "audio_speech": false
    },
    "name": "mistral-medium-2508",
    "description": "Update on Mistral Medium 3 with improved capabilities.",
    "max_context_length": 131072,
    "aliases": [
      "mistral-medium-latest",
      "mistral-medium",
      "mistral-vibe-cli-with-tools"
    ],
    "deprecation": null,
    "deprecation_replacement_model": null,
    "default_model_temperature": 0.3,
    "type": "base"
  },

So that has the alias "mistral-medium-latest", but the official ID is "mistral-medium-2508" which suggests it's the model they released in August 2025.

But... that 1777479384 timestamp decodes to Wednesday, April 29, 2026 at 04:16:24 PM UTC

So is that the new Mistral Medium?

simonw2• 3 months ago

Some poking around in the source code for https://github.com/mistralai/mistral-vibe got me to this:

  curl https://api.mistral.ai/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $(llm keys get mistral)" \
  -d '{
    "model": "mistral-medium-3.5",
    "messages": [
      {"role": "user", "content": "Generate an SVG of a pelican riding a bicycle"}
    ]
  }'

Which did work: https://gist.github.com/simonw/f3158919b18d2c47863b0a5dc257a... - it's pretty disappointing.

Weird that it doesn't show up in the model list:

  curl https://api.mistral.ai/v1/models \
    -H "Content-Type: application/json" \
    -H "Authorization: Bearer $(llm keys get mistral)" | jq

Mashimo5• 3 months ago

I also did some SVG tests, it's really bad.

https://chat.mistral.ai/chat/897fbe7d-b1ae-4109-9b29-f3ccc4f...

minimaxir• 3 months ago

It's funny that 128B is now considered Medium. I remember back in the day when 355M parameters was considered medium with GPT-2.

speedgoose• 3 months ago

And GPT-2 1.5B was considered too dangerous to release.

They were perhaps right.

Matl5• 3 months ago

considered that by OpenAI for marketing purposes that is

But yes, perhaps it would have been better for all of us if they haven't.

Giorgi2• 3 months ago

Oh they are still a thing?! Completely forgot about Mistral. I am assuming they are still burning trough investor money.

sev_verso• 3 months ago

What's better than Voxtral for locally processed voice input? More competition is always better.

danelski2• 3 months ago

I believe they'll get profitable sooner than their frontier competition. Their operating costs seem to peanuts compared to the providers they're compared to most often while having the local advantage of not being Chinese nor American.

kergonath• 3 months ago

they are still burning trough investor money

Difficult to say, this information is not really public. That said, those investors include EU agencies and European multinational companies and governments. It’s not as flashy as the ridiculous sums OpenAI is getting but it should be enough to keep them going for a while.

They also have a different business model. They are selling their expertise to fine tune and adapt their models to on-premises computers (which they can help you build) to handle confidential data and information. I would not be surprised that the revenue they get from normal people is negligible in comparison.

Giorgi2• 3 months ago

Ooh, ok so people got all worked up because it is EU vs USA thing.

Havoc2• 3 months ago

Think they’re positioned pretty well. They’ve got an edge in the European corporate space and don’t have ungodly large numbers to hit

simjnd15• 3 months ago

I'm not sure what people are on in the comments. It doesn't beat the other models, but it sure competes despite its size.

GLM 5.1 is an excellent model, but even at Q4 you're looking at ~400GB. Kimi K2.5 is really good too, and at Q4 quantization you're looking at almost ~600GB.

This model? You can run it at Q4 with 70GB of VRAM. This is approaching consumer level territory (you can get a Mac Studio with 128GB of RAM for ~3500 USD).

For the Claude-pilled people, I don't know if you only run Opus but when I was on the Pro plan Sonnet was already extremely capable. This beats the latest Sonnet while running locally, without anyone charging you extra for having HERMES.md in your repo, or locking you out of your account on a whim.

Mistral has never been competitive at the frontier, but maybe that is not what we need from them. Having Pareto models that get you 80% of the frontier at 20% of the cost/size sounds really good to me.

redrove• 3 months ago

It’s 128b dense model. Good luck getting more than 3t/s out of a mac. It doesn’t matter if it fits or not.

zozbot2349• 3 months ago

You could run it on a single Mac Studio with M3 Ultra, or two Mac Studios with M4 Max at higher perf than that. And lightly quantizing this could give us modern dense models in the ~80GB size range, which is a very compelling target.

gregsadetsky• 3 months ago

I didn't know about HERMES.md ... (??) - found information here for others who are curious https://github.com/anthropics/claude-code/issues/53262

giancarlostoro2• 3 months ago

That is insane, if you billed me an extra $200 for a bug in your system I'd flat out cancel my subscription. If you're not going to credit that back to me, you don't deserve anymore of my money. I'm a Claude first guy, but if you're going to bill me incorrectly, that's on you, own it, fix it.

gnulinux• 3 months ago

This github thread is incredible, thanks for sharing. This link should be its own HN topic.

YetAnotherNick• 3 months ago

It has similar SWE bench score to qwen 3.6 27b[1]. No one is comparing it to frontier.

[1]: There is no other common benchmark in the blog.

simjnd15• 3 months ago

That's more a testament of how good Qwen3.6 27B is (it really is great) more than how bad this one is IMO. Gemma 4 31B was already good, but Qwen3.6 27B is incredible for its size.

freakynit3• 3 months ago

I was hoping a lot from it... but this one, is not up to that mark. For example, here is it's comparion with 4.7x smaller model, qwen3.7-27b.

https://chatgpt.com/share/69f239e8-7414-83a8-8fdd-6308906e5f...

Tldr: qwen3.6-27b, a 4.7x smaller model, have similar performance.

r0b052• 3 months ago

That's a chatgpt summary. Actual usage would a better test.

lostmsu• 3 months ago

To be fair MoE from Qwen itself had the same "problem". 3.5 122B MoE was same or worse than 3.5 27B. Yet to see 122B 3.6.

UPD. NVM, Mistral Medium 3.5 is dense. So yes, it is worse in every way.

Aurornis3• 3 months ago

This model? You can run it at Q4 with 70GB of VRAM. This is approaching consumer level territory (you can get a Mac Studio with 128GB of RAM for ~3500 USD).

The one thing I would want everyone curious about local LLMs to know is that being able to run a model and being able to run a model fast are two very different thresholds. You can get these models to run on a 128GB Mac, but we need to first tell if Q4 retains enough quality (models have different sensitivities to quantization) and how fast it runs.

For running async work and background tasks the prompt processing and token generation speeds matter less, but a lot of Mac Studio buyers have discovered the hard way that it's not going to be as responsive as working with a model hosted in the cloud on proper hardware.

For most people without hard requirements for on-site processing, the best use case for this model would be going through one of the OpenRouter hosted providers for it and paying by token.

This beats the latest Sonnet while running locally

Almost every open weight model launch this year has come with claims that it matches or exceeds Sonnet. I've been trying a lot of them and I have yet to see it in practice, even when the benchmarks show a clear lead.

zozbot2349• 3 months ago

Cloud hardware is not inherently more "proper" than what's being proposed here, there's nothing wrong per se about targeting slower inference speeds in an on prem single-user context.

simjnd15• 3 months ago

The one thing I would want everyone curious about local LLMs to know is that being able to run a model and being able to run a model fast are two very different thresholds. You can get these models to run on a 128GB Mac, but we need to first tell if Q4 retains enough quality (models have different sensitivities to quantization) and how fast it runs.

Very valid. This is an active area of research, and there are a lot of options to try out already today.

- People have successfully used TurboQuant to quantize model weights (TQ3_4S), not just the context KV, to achieve smaller sizes than Q4 (~3.5 bpw) with much better PPL and faster decoding.

- Importance-weighted quantization (e.g. IQ4) also provides way better PPL, KDL, etc. at the same size as a Q4.

- DFlash (block diffusion for speculative decoding) needs a good drafting model compatible with the big model, but can provide an uplift up to 5x in decoding (although usually in the 2-2.5x range)

- Forcing a model's thinking to obey a simple grammar has been shown to improve results with drastically lower thinking output (faster effective result generation) although that has been more impactful on smaller models.

We should be skeptical, but it's definitely trending in the right direction and I wouldn't be surprised if we are indeed able to run it at acceptable speeds.

Almost every open weight model launch this year has come with claims that it matches or exceeds Sonnet. I've been trying a lot of them and I have yet to see it in practice, even when the benchmarks show a clear lead.

This hasn't been my experience. After Anthropic's started their shenanigans I've switched to exclusively using open-weights models via OpenRouter and OpenCode and I can't really tell a difference (for better or for worse).

Computer0• 3 months ago

Sure but for a casual conversational use case I have not found speed to be a huge barrier. I chatted with a 100b model using ddr5 only on a plane recently and it was fine. It's mainly that I cannot do data classification and coding tasks in a timely manner.

nijave2• 3 months ago

Almost every open weight model launch this year has come with claims that it matches or exceeds Sonnet. I've been trying a lot of them and I have yet to see it in practice, even when the benchmarks show a clear lead.

This has been my experience as well. I've been testing an agent built with Strands Agents which receives a load balancer latency alert and is expected to query logs with AWS Athena (Trino) then drill down with Datadog spans/traces to find the root cause. Admittedly, "devops" domain knowledge is important here

My notes so far:

"us.anthropic.claude-sonnet-4-6" # working, good results

"us.anthropic.claude-sonnet-4-20250514-v1:0" # has problems following the prompt instructions

"us.anthropic.claude-sonnet-4-5-20250929-v1:0" # working, good results

"us.anthropic.claude-opus-4-5-20251101-v1:0"

"us.anthropic.claude-opus-4-6-v1" # best results, slower, more expensive

"amazon.nova-pro-v1:0" # completely fails

"openai.gpt-oss-120b-1:0" # tool calling broken

"zai.glm-5" # seems to work pretty well, a little slow, more expensive than Sonnet

"minimax.minimax-m2.5" # didn't diagnose correctly

"zai.glm-4.7" # good results but high tool call count, more expensive than Sonnet

"mistral.mistral-large-3-675b-instruct" # misdiagnosed--somehow claimed a Prometheus scrape issue was involved

"moonshotai.kimi-k2.5" # identified the right endpoints but interpreted trace data/root cause incorrectly

"moonshot.kimi-k2-thinking" # identified endpoint, 1 correct root cause, 1 missing index hallucination

Using models on AWS Bedrock. I let Claude Code w/ Opus 4.7 iterate over the agent prompt but didn't try to optimize per model. Really the only thing that came close to Sonnet 4.5 was GLM-5. The real kicker is, Sonnet is also the cheapest since it supports prompt caching

The Kimi ones were close to working but didn't quite make the mark

parsimo20102• 3 months ago

being able to run a model and being able to run a model fast are two very different thresholds

Specifically speaking, on my Strix Halo machine with (theoretical) memory bandwidth of 256 GB/s, a 70 GB model can't generate faster than 256/70= 3.65 t/s. The logic here is that a dense model must do a full read of the weights for each token. So even if the GPU can keep up, the memory bandwidth is limiting.

A Mac M5 Pro is faster with a bandwidth of 307 GB/s, but that's only a little faster.

This thing is going to be slow on consumer hardware. Maybe that is useful for someone, but I probably prefer a faster model in most cases even if the model isn't quite as smart. Qwen3.6 35B-A3B generates about 50 t/s on my machine, so it can make mistakes, be corrected, and try again in the same time that this model would still be thinking about its first response.

notatoad• 3 months ago

being able to run a model fast is definitely more useful, but being able to run a model slowly for free is still super useful. agentic workflows are maturing all the time.

yes, if i'm directly interacting with the LLM, i want it to be reasonably fast. but lately i've been queueing up a bunch of things when i go for lunch, or leaving things running when i go home at the end of the day. and claude doesn't keep working on that all night, it runs for an hour or so, gets to a point where it needs more input from me, and gives me some stuff to review in the morning. that could run 16x slower and still be just as useful for me.

2ndorderthought8• 3 months ago

The point is it's open weight and is tiny compared to a lot of it's competitors. 4gpus for world class performance - sweet!

DeathArrow2• 3 months ago

This model? You can run it at Q4 with 70GB of VRAM. This beats the latest Sonnet while running locally

Not sure it will beat Sonet at Q4.

This is approaching consumer level territory (you can get a Mac Studio with 128GB of RAM for ~3500 USD).

For $3500 I can get 7-8 years of GLM using coding plans, have a faster model and much better code quality.

kobalsky• 3 months ago

For $3500 I can get 7-8 years of GLM

mind sharing where's the go to place to pay for open models?

simjnd15• 3 months ago

Not sure it will beat Sonet at Q4.

Very valid. Importance-weighted quantization and TurboQuant on model weights can reduce loss a lot compared to "traditional" Q4 so one can be hopeful.

For $3500 I can get 7-8 years of GLM using coding plans, have a faster model and much better code quality

But you will own no computer, and that's also assuming prices stay what they are. Anyway my point was not whether or not it makes financial sense for everyone. A lot of people are very happy not owning their movies, software, games, cars or house. I'm just happy there is a future where the people can own and locally run the tech that was trained on their stolen data.

Ancapistani• 3 months ago

For $3500 I can get 7-8 years of GLM using coding plans, have a faster model and much better code quality.

I know HN's distaste for crypto, but I do my inference (for personal stuff - not my employer) through Venice. I was in the airdrop for VVV, and kept as much of it staked as I could. I have ~$40/day in inference as long as that service lasts.

These days the multiplier is about 1000x last I checked; if you want $10/day in inference and can lock up $10k in VVV, you get ~$10/day in inference plus (currently) ~16% APY in the form of more VVV.

I'm not sure I'd want to invest that much if I had to today, but it's a reasonable option. The risk of VVV going to $0 seems pretty small to me.

liuliu• 3 months ago

The competition is on DeepSeek v4 Flash for similar size / deployment target.

simjnd15• 3 months ago

DeepSeek v4 Flash is still over 100GB at Q4 IIRC, and Q4 has generally been the sweet spot. Although it's an MoE so it might run a lot faster that this dense Mistral model if you have the RAM.

giancarlostoro2• 3 months ago

For the Claude-pilled people, I don't know if you only run Opus but when I was on the Pro plan Sonnet was already extremely capable.

Before February I was able to use Opus on High exclusively on my Max plan no problem. Now I've shifted to just using Sonnet on high and yeah, its pretty capable. I love that, Claude Pilled. ;)

simjnd15• 3 months ago

Yeah I love Claude, amazing models. Anthropic has very quickly burned most of the goodwill I had for it so I still ended up cancelling my subscription.

UncleOxidant• 3 months ago

Yeah, you can run it locally if you have enough VRAM, but the reports trickling in are saying about 3 tok/sec. This was on a Strix Halo box which definitely has the needed VRAM, but isn't going to have as high mem bandwidth as a GPU card, it's going to be similar on a Mac - that's the dilemma... the unified memory machines have the VRAM, but the bandwidth isn't great for running dense models. This size of a dense model is only going to be runnable (usefully) by very few people who have multiple GPU cards with enough memory to add up to about 70GB.

simjnd15• 3 months ago

I don't think this is quite correct, a Strix Halo box usually has 256 GB/s memory bandwidth. An M5 Max has 614 GB/s. An M3 Ultra (no M4 or M5 Ultra) has 820 GB/s. It's still not GDDR or HBM territory, but still significantly faster.

That's the edge of Apple Silicon for AI. When they scale up the chip they add more memory controllers which adds more channels and more bandwidth.

But yeah in the end it's still going to be only a handful of people that can run it.

What I meant is that I think researching and developing smaller more powerful model is more interesting than chasing the next 3T parameter model while burning through VC money and squeezing your customer base more and more aggressively.

zackangelo• 3 months ago

Isn't Kimi K2.6 natively INT4?

simjnd15• 3 months ago

I don't think any models are natively INT4? I wouldn't see the point to nerf the model out-of-the-box.

deepsquirrelnet• 3 months ago

I would love to be able to run frontier locally, but I think the larger importance of open weight models is price accountability.

In the US with our broken system of capitalism, it’s the only way we can tether these companies to reality. Left to their own devices, I’m not convinced they would actually compete with each other on price.

Buy nobody like to talk about how “moat” building is fundamentally anti-competitive, even in name.

Funny that self proclaimed capitalists hate the system in practice. Commodity pricing is what truly terrifies them.

simjnd15• 3 months ago

I'm not necessarily interested in having frontier locally. You don't need to be frontier to be a very good and useful coding agent. I agree with your point on price accountability though. Hopefully no tariff comes down on the Chinese and European open-weight models.

ksubedi• 3 months ago

Let's not forget Qwen 35B A3B MoE. It gets better performance than this in all the metrics for a fraction of the memory / compute footprint.

Sad to see all the non Chinese open source models being at least one generation behind.

simjnd15• 3 months ago

Qwen3.6 27B is even more impressive IMO. Dense so it doesn't run as fast but it's so good.

revolvingthrow• 3 months ago

Eh. Those results would be noteworthy if it was a a MoE. A 120B dense? Firmly in meh territory.

gregorygoc• 3 months ago

Why do you care?

WhitneyLand• 3 months ago

“This beats the latest Sonnet while running locally”

Not really.

- The benchmarks are based on F8_E4M3 and you’re not running that on any Mac.

- Sonnet has a 1M token context window. This is 256k but again you’re probably not even getting that locally.

- Sonnet is fast over the wire. This is going to be much slower.

trueno2• 3 months ago

the benchmarks we're using to measure llm's do no justice when everyone's mental-benchmark is simply "is it going to feel like using claude" and the answer is still no. the entire llm space is stuffed with tons of crazy datapoints and vernacular that barely paint the picture of the mental benchmark everyone is after.

i too am desperate to just sever ties with these big providers, my fingers are crossed we get there within the constraints of local hardware even if that means me spending 3-5k i just want off this wild ride.

trvz• 3 months ago

Sonnet is fast over the wire.

Except when it’s unavailable. For sovereignity, the downsides are worth it to some.

varispeed2• 3 months ago

Not sure if 1M token window is meaningful with Sonnet/Opus. The models go dumb quickly as context increases making them unusable (that is if you get routed to actual Opus, otherwise they are just dumb regardless of context window).

varispeed2• 3 months ago

It doesn't beat the other models, but it sure competes despite its size.

But what is the rationale for running a dumb model? Because it can ocasionally produce something passable?

I don't get where is the value apart from mild entertainment, as in "I am somewhat of Anthropic myself".

simjnd15• 3 months ago

Are you dumb because you're not Einstein? Intelligence is a spectrum. Just because you're not #1 doesn't mean you're dumb. A lot of small models are not frontier but are still very competent and are very useful coding agent. It may take better prompting and more guiding, but that can be a reasonable tradeoff for some people.

Mashimo5• 3 months ago

Compared to all other hosted LLMs that I have tested, Mistral seems to be the only one with rather strict CSP headers. When you ask them to create a website with some javascript library it will not preview, even though le chat offers canvas mode.

Sometimes when a new release comes around from any provider I just want to test it a bit on the web. without paying and using an agent harness.

Why are they like this ;_;

Edit: Christ on a bike it's bad at drawing SVGs https://chat.mistral.ai/chat/23214adb-5530-4af9-bb47-90f5219...

2ndorderthought8• 3 months ago

I have never wanted, needed or hoped to draw svgs with an LLM. All of the models suck at it, some are just more fun or something.

Mashimo5• 3 months ago

I can't speak for what you consider sucking, but there is a significant difference between Mistral and Kimi or Gemini. I find the others to be usable for my needs.

andai• 3 months ago

Claude volunteered this the other day:

https://iili.io/BsfyNXR.jpg

(I think the hair was unintentional, but it is impossible to be sure.)

SyneRyder• 3 months ago

Edit: Christ on a bike it's bad at drawing SVGs

On the bike would be an improvement. Geez.

I know SVGs may not be the best benchmark, but that matches my experience of trying to run a (previous) Mistral model in Mistral Vibe, asking it to help me configure an MCP server in Vibe. It confidently explained that MCP is the MineCraft Protocol and then began a search of my computer looking for Minecraft binaries.

ffsm8• 3 months ago

Are those cherries overlayed ontop of boobs with a bus to the side, driving towards a rock? Scnr

seb_lz• 3 months ago

I'm using mistral-medium-2508 for some text transformation operations. It's giving me better results than mistral-large for my use cases. Looking forward to testing this new model, although I'm not sure if it's really meant at replacing the previous medium model since it's a lot more expensive and presented more as a coding / agentic model (mistral-medium-2508 was priced $0.4/$2 per 1M tokens, mistral-medium-3.5 is $1,5/$7.5).

hulk-konen2• 3 months ago

I actually use Mistral Large to go through some large text chunks (in production). It gives about the same level of results as Sonnet, while being 90% cheaper. Definitely wouldn't use it for coding, but for this text-analyzing task it has been great. Much better than all the latest Chinese models, for example.

So I was waiting for this release and it's... 5x more expensive than the latest Mistral Large. So now I'm worried they'll pull the plug on the cheap Large when their releases roll over to that one.

zozbot2349• 3 months ago

Why does this matter if the model is open? It can be offered by competitive third-party providers, there's no rug pull.

barrell• 3 months ago

Yeah I use Mistral Large for a lot of formatting work. For this one use case of mine, it outperforms frontier models by a significant margin. I've found tons of use cases for mistral small as well.

I'd love to use Mistral for more tasks, but Mistral Large doesn't quite cut it for all tasks. So on the one hand, I'm excited there is another model, and presumably more performant based on the price? But the fact it's a "Medium" and 5x the price of the Large definitely concerns me.

The entire release is also about Vibe Coding, and so I'm not even sure if this model is applicable outside of coding, or even worth testing.

syntaxing2• 3 months ago

This is a very interesting strategy that might pay off. This model is a very good option for enterprise self host. I would argue a lot of companies are VRAM constrained rather than compute constrained. You could fit 4-5 running instances on one H100 cluster where you can only fit 1-2 Kimi K2 or GLM5.

2001zhaozhao• 3 months ago

This is 128B dense though. the K/V cache on long context is going to be massive

Havoc2• 3 months ago

Don’t think kv size correlates to dense/moe

syntaxing2• 3 months ago

With turbo quant, you would reduce it by over 6X.

maelito• 3 months ago

Given what Vibe already did in the previous versions with codestral-v2, that's great news. Keep up the good work ! I don't want to depend on the world's two hungry superpowers.

Alifatisk• 3 months ago

A 1000B model, can we call it 1KB model?

schipperai2• 3 months ago

With most OSS releases being MoEs, and modern GPUs optimized for MoEs, can somebody with knowledge of the topic explain or speculate why Mistral might have opted for a dense model?

ac29• 3 months ago

Modern GPUs aren't optimized for MoEs though?

The advantage to a dense model like this Mistral one is that it is as smart as a much larger MoE model so it can fit on less GPUs. The tradeoff is that it is much slower since it has to read 100% of its weights for every token, MoE models typically only read about a tenth (though sparsity levels vary).

schipperai2• 3 months ago

Thanks, makes sense. I meant Blackwell is explicitly optimized for MoEs.

andhuman• 3 months ago

The Vibe CLI is really bad on Windows, sure they don’t officially support it, so can’t blame them, but a FYI for anyone wanting to try it. It can’t get find and replace right.

KronisLV2• 3 months ago

For it's size, that's really good! Though I bet it being a dense model probably helps a lot, if it was MoE at that size, I bet the benchmark performance would go quite a bit down (which consequently would also mean that I'd at least be able to run it with decent tokens/second, with the bunch of Nvidia L4 cards available to me, which presently are only okay with MoE models).

It's cool that they added comparisons to their own Mistral Small 4 119B A7B, which kind of shows that! They could have also included comparisons to something like Qwen Coder Next 80B A3B (or maybe the newer Qwen 3.6 35B A3B, or the 27B dense one), maybe DeepSeek V4 Flash 284B A13B, or the older GPT-OSS 120B A5B to illustrate that difference and where their model sits even better, it would probably give a more positive picture than just comparing themselves against a bunch of bigger models!

Come to think of it, alongside throwing some money at DeepSeek not just Anthropic, I probably should get a Mistral subscription as well sometime, to see how they perform on various tasks - cause they seem pretty cost effective and it's nice to support at least some EU orgs: https://mistral.ai/pricing

jollymonATX• 3 months ago

Ouch. Maybe they have a captive buying market to insulate them from actual market forces or ???

antirez2• 3 months ago

The problem with this model is that DeepSeek v4 Flash runs quite well quantized to 2 bit (see https://github.com/antirez/llama.cpp-deepseek-v4-flash), at 30 t/s generation and 400 t/s prefill in a M3 Ultra (and not too much slower on a 128GB MacBook Pro M3 Max). It works as a good coding agent with opencode/pi, tool calling is very reliable, and so forth. All this at a speed that a 120B dense model can never achieve. So it has to compete not just with models that fit 4-bit quantized the same size, but with an 86GB GGUF file of DeepSeek v4 Flash, and it is not very easy to win in practical terms for local inference.

Note: I have more uncommitted speed improvements in my tree that I'll push soon, the current tree could be a little bit slower but not much, still super usable.

I don't understand one thing about Mistral, which I'm a fan being in Europe: they opened the open weights MoE show with Mixtral. Why are they now releasing dense models of significant sizes? In this way you don't compete in any credible space, nor local inference, nor remote inference since the model is far from SOTA and not cheap to serve. So why they are training such dense big models? Dense models have a place in the few tens of billion parameters, as Qwen 3.6 27B shows, but if you go 5 times that, it is no longer a fit, unless you are crushing with capabilities anything requiring the same VRAM, which is not the case.

zozbot2349• 3 months ago

Your GitHub link only says "The model quantized in this way behaves very very well in the chat, frontier-model vibes, but it was not extensively tested." This is hardly relevant to how it behaves in agentic workflows, we're aware of how often they degrade severely with Q2 quantization. If this quantized Flash can keep up reasonable quality and performance at larger context lengths (which seems to be a key feature of the V4 series) it could be a very reasonable competitor to models in the same weight class like Qwen 3 Coder-Next 80B.

antirez2• 3 months ago

Nope it works great with opencode as a agent, you can build a game or things like that. It works. The trick is a mix among the quantization I used, which is very asymmetric, and the fact that I guess DeepSeek v4 Flash tolerates extreme quantizations better than anything I saw in the past.

What I used was up/gate of routed models, IQ2_XXS, out -> Q2_K, then I quantized routing, projections, shared experts to Q8. The trick is that the very sensible parts are a small amount of the weights, and they are kept very high quality.

vjay15• 3 months ago

Mistral is playing the long game here ngl, lower sized models, lower costs, overall good enough performance!

dznan• 3 months ago

I hack you

Scroll_Swe• 3 months ago

Le Chat fast is really good now. Did a few normal queries and noticed and improvement. Very well done.

Mistral Medium 3.5

Related Stories