▲ 70 points
back
20 comments
It's not cheaper than Deepseek V3.1, though, and Deepseek outperforms on nearly everything. And only between 1- 3x the throughput based on the openrouter metrics (near equivalent throughput if you use an FP8 quant). Wish I could be a little more excited about this one.
Mistral seems to be the only company that doesn‘t fake benchmarks. That makes it of course not exciting. But it doesn‘t have to be to be useful.
What are you basing this on?
Love the information on this guy's website, dislike the layout of information and use of Google Sheets on pages. Makes it difficult to navigate and grok.
I guess this one (Mistral Medium 3) won't be open?
They are dropping some hints about a larger model and something relating to open models in the final paragraph:
> With the launches of Mistral Small in March and Mistral Medium today, it’s no secret that we’re working on something ‘large’ over the next few weeks. With even our medium-sized model being resoundingly better than flagship open source models such as Llama 4 Maverick, we’re excited to ‘open’ up what’s to come :)
Hmm, bearish on Mistral now. I thought they had a plan to monetize that was much more sophisticated than "freemium".
There's no reason for someone to use this over Anthropic, OpenAI, etc ... they don't even outperform them.
Business can self deploy Mistral models in their infra.
You cant do that with the providers you listed.
Indeed, however in practice almost nobody does it.
They event hint at it on their PR:
"Mistral La Plateforme and Amazon Sagemaker, and soon on IBM WatsonX, NVIDIA NIM, Azure AI Foundry, and Google Cloud Vertex"
My point being, why would you pay to use a Mistral model hosted on Azure, instead of using any other company model hosted on Azure?
My answer to that, yesterday, would have been "because the model is free and unrestricted, I only pay for hardware"; that premise is gone today.
> Indeed, however in practice almost nobody does it.
They released it at the same time of the article we are discussing on...
The big dogs also offer on-prem if you are important enough. Here's article for Gemini: https://cloud.google.com/blog/products/ai-machine-learning/r...
EU Legal / Privacy teams are generally fine with OpenAI models running on Azure from my experience (I assume for GDPR reasons). That's "my infra" enough.
And you get decent quality models without the hassle of maintaining the infra.
Ah, but here's the thing!
If you carefully read their performance chart, they beat every open source model they bothered to list.
So if we were in a world in which they released this model, and only the listed open source models existed, they would be le SOTA!
DeepSeek wins most benchmarks according to that chart, so not quite SOTA, and there's no mention of model size so it's hard to compare efficiency
> There's no reason for someone to use this over Anthropic, OpenAI, etc
Uh except that you can host this yourself? You just have to license the model. What is the issue?
>You just have to license the model.
That's the issue.
The medium model is crazy fast. Reminds me of maverick on groq (except better according to their own testing).
The only hint at its size is that it requires "self-hosted environments of four GPUs and above".