I'm sure I'm not the only one that thinks "Vibe CLI" sounds like an unserious tool. I use Claude Code a lot and little of it is what I would consider Vibe Coding.
Look interesting, eager to play around with it! Devstral was a neat model when it released and one of the better ones to run locally for agentic coding. Nowadays I mostly use GPT-OSS-120b for this, so gonna be interesting to see if Devstral 2 can replace it.
I'm a bit saddened by the name of the CLI tool, which to me implies the intended usage. "Vibe-coding" is a fun exercise to realize where models go wrong, but for professional work where you need tight control over the quality, you can obviously not vibe your way to excellency, hard reviews are required, so not "vibe coding" which is all about unreviewed code and just going with whatever the LLM outputs.
But regardless of that, it seems like everyone and their mother is aiming to fuel the vibe coding frenzy. But where are the professional tools, meant to be used for people who don't want to do vibe-coding, but be heavily assisted by LLMs? Something that is meant to augment the human intellect, not replace it? All the agents seem to focus on off-handing work to vibe-coding agents, while what I want is something even tighter integrated with my tools so I can continue delivering high quality code I know and control. Where are those tools? None of the existing coding agents apparently aim for this...
Does anyone know where their SWE-bench Verified results are from? I can't find matching results on the leaderboards for their models or the Claude models and they don't provide any links.
I am very disappointed they don't have an equivalent subscription for coding to the 200 EUR ChatGPT or Claude one, and it is only available for Enterprise deployments.
The only thing I found is a pay-as-you-go API, but I wonder if it is any good (and cost-effective) vs Claude et al.
> Devstral 2 ships under a modified MIT license, while Devstral Small 2 uses Apache 2.0. Both are open-source and permissively licensed to accelerate distributed intelligence.
Uh, the "Modified MIT license" here[0] for Devstral 2 doesn't look particularly permissively licensed (or open-source):
> 2. You are not authorized to exercise any rights under this license if the global consolidated monthly revenue of your company (or that of your employer) exceeds $20 million (or its equivalent in another currency) for the preceding month. This restriction in (b) applies to the Model and any derivatives, modifications, or combined works based on it, whether provided by Mistral AI or by a third party. You may contact Mistral AI (sales@mistral.ai) to request a commercial license, which Mistral AI may grant you at its sole discretion, or choose to use the Model on Mistral AI's hosted services available at https://mistral.ai/.
Less than a year behind the SOTA, faster, and cheaper. I think Mistral is mounting a good recovery. I would not use it yet since it is not the best along any dimension that matters to me (I'm not EU-bound) but it is catching up. I think its closed source competitors are Haiku 4.5 and Gemini 3 Pro Fast (TBA) and whatever ridiculously-named light model OpenAI offers today (GPT 5.1 Codex Max Extra High Fast?)
I just end up using most of these models with Claude Code as the tooling because it just seems to work better than anything else. Crush also works well.
offtopic but it hurts my eyes: I dislike for their font choice and their "cool looks" in their graphics.
Surprising and good is only: Everything including graphics fixed when clicking my "speedreader" button in Brave. So they are doing that "cool look" by CSS.
Very nice that there's a coding cli finally. I have a Mistral Pro account. I hope that it will be included. It's the main reason to have a Pro account tbh.
Let's see which company becomes the first to sell "coding appliances": hardware with a model good enough for normal coding.
If Mistral is so permissive they could be the first ones, provided that hardware is then fast/cheap/efficient enough to create a small box that can be placed in an office.
My Macbook Pro with an M4 Pro chip can handle a number of these models (I think it has 16GB of VRAM) with reasonable performance, my bottleneck continuously is the token caps. I assume someone with a much more powerful Mac Studio could run way more than I can, considering they get access to about 96GB of VRAM out of the system RAM iirc.
I was briefly excited when Mistral Vibe launched and mentions "0 MCP Servers" in its startup screen... but I can't find how to configure any MCP servers. It doesn't respond to the /mcp command, and asking Devstral 2 for help, it thinks MCP is "Model Context Preservation". I'd really like to be able to run my local MCP tools that I wrote in Golang.
I'm team Anthropic with Claude Max & Claude Code, but I'm still excited to see Mistral trying this. Mistral has occasionally saved the day for me when Claude refused an innocuous request, and it's good to have alternatives... even if Mistral / Devstral seems to be far behind the quality of Claude.
Let's say you had a hardware budget of $5,000. What machine would you buy or build to run Devstral Small 2? The HuggingFace page claims it can run on a Mac with 32 GB of memory or an RTX 4090. What kind of tokens per second would you get on each? What about DGX Spark? What about RTX 5090 or Pro series? What about external GPUs on Oculink with a mini PC?
dual 3090's (24GB each) on 8x+8x pcie has been a really reliable setup for me (with nvlink bridge... even though it's relatively low bandwidth compared to tesla nvlink, it's better than going over pcie!)
48GB of vram and lots of cuda cores, hard to beat this value atm.
If you want to go even further, you can get an 8x V100 32GB server complete with 512GB ram and nvlink switching for $7000 USD from unixsurplus (ebay.com/itm/146589457908) which can run even bigger models and with healthy throughput. You would need 240V power to run that in a home lab environment though.
Looks like another Deepseek distil like the new Ministrals. For every other use case that would be an insult, but for coding that's a great approach given how much lead in coding performance Qwen and Deepseek have on Mistral's internal datasets. The Small 24B seems to have a decent edge on 30BA3B, though it'll be comparatively extremely slow to run.
I gave it the job of modifying a fairly simple regex replacement and it took a while over 5 minutes, claude failed on the same prompt (which surprised me), codex did a similar job but faster. So all in all not bad!
Open sourcing the TUI is pretty big news actually. Unless I missed something, I had to dig a bit to find it, but I think this is it: https://github.com/mistralai/mistral-vibe
62 comments
[ 2.9 ms ] story [ 78.3 ms ] threadEven the Gemini 3 announcement page had some bit like "best model for vibe coding".
I'm a bit saddened by the name of the CLI tool, which to me implies the intended usage. "Vibe-coding" is a fun exercise to realize where models go wrong, but for professional work where you need tight control over the quality, you can obviously not vibe your way to excellency, hard reviews are required, so not "vibe coding" which is all about unreviewed code and just going with whatever the LLM outputs.
But regardless of that, it seems like everyone and their mother is aiming to fuel the vibe coding frenzy. But where are the professional tools, meant to be used for people who don't want to do vibe-coding, but be heavily assisted by LLMs? Something that is meant to augment the human intellect, not replace it? All the agents seem to focus on off-handing work to vibe-coding agents, while what I want is something even tighter integrated with my tools so I can continue delivering high quality code I know and control. Where are those tools? None of the existing coding agents apparently aim for this...
What matters is high quality specifications including test cases
The only thing I found is a pay-as-you-go API, but I wonder if it is any good (and cost-effective) vs Claude et al.
As long as it doesn't mean 10x worse performance, that's a good selling point.
Uh, the "Modified MIT license" here[0] for Devstral 2 doesn't look particularly permissively licensed (or open-source):
> 2. You are not authorized to exercise any rights under this license if the global consolidated monthly revenue of your company (or that of your employer) exceeds $20 million (or its equivalent in another currency) for the preceding month. This restriction in (b) applies to the Model and any derivatives, modifications, or combined works based on it, whether provided by Mistral AI or by a third party. You may contact Mistral AI (sales@mistral.ai) to request a commercial license, which Mistral AI may grant you at its sole discretion, or choose to use the Model on Mistral AI's hosted services available at https://mistral.ai/.
[0] https://huggingface.co/mistralai/Devstral-2-123B-Instruct-25...
How is that a measure of model size? It should either be parameter size, activated parameters, or cost per output token.
Looks like a typo because the models line up with reported param sizes.
Why does every AI provider need to have its own tool, instead of contributing to existing tools like Roo Code or Opencode?
Pretty good for a 123B model!
(That said I'm not 100% certain I guessed the correct model ID, I asked Mistral here: https://x.com/simonw/status/1998435424847675429)
Surprising and good is only: Everything including graphics fixed when clicking my "speedreader" button in Brave. So they are doing that "cool look" by CSS.
If Mistral is so permissive they could be the first ones, provided that hardware is then fast/cheap/efficient enough to create a small box that can be placed in an office.
Maybe in 5 years.
I'm team Anthropic with Claude Max & Claude Code, but I'm still excited to see Mistral trying this. Mistral has occasionally saved the day for me when Claude refused an innocuous request, and it's good to have alternatives... even if Mistral / Devstral seems to be far behind the quality of Claude.
48GB of vram and lots of cuda cores, hard to beat this value atm.
If you want to go even further, you can get an 8x V100 32GB server complete with 512GB ram and nvlink switching for $7000 USD from unixsurplus (ebay.com/itm/146589457908) which can run even bigger models and with healthy throughput. You would need 240V power to run that in a home lab environment though.
Just call it Mistral License & flush it down
After querying the model about .NET, it seems that its knowledge comes from around June 2024.
Going to start hacking on this ASAP