Mistral ships 128B dense model
- Mistral released Mistral Medium 3.5 in public preview on April 29, folding reasoning, coding, vision, and chat into one open-weight 128B model. - The telling detail is deployment: Mistral says the dense model has a 256K context window and can self-host on as few as four GPUs. - That matters because open models usually trade capability for runnability; Mistral is trying to narrow that gap without closing the weights.
Open-weight AI models usually force a tradeoff. You get control and local deployment, but you give up some of the polish and breadth of the best hosted systems. Mistral is trying to break that pattern. On April 29, it shipped Mistral Medium 3.5 — a 128B dense model that rolls instruction-following, reasoning, coding, and vision into one set of weights, while still claiming it can run self-hosted on as few as four GPUs. (mistral.ai) ### What actually shipped? This is Mistral Medium 3.5, released in public preview and made available as open weights on Hugging Face under Mistral’s modified MIT license. The model is dense, not mixture-of-experts, supports a 256K context window, takes text and image input, and lets developers toggle “reasoning effort” per request instead of swapping between separate models. (mistral.ai) ### Why does “dense 128B” matter? Dense means the whole model is engaged on each token, not just a subset of experts. That usually makes the system simpler to reason about operationally, but it also sounds expensive. So the surprise here is not just the size. It’s that Mistral is pitching a frontier-ish dense model as something practical enough for smaller self-hosted setups, rather than a cloud-only showcase. (mistral.ai) ### What problem is Mistral trying to solve? Model sprawl. Mistral had separate lines for chat, reasoning, and coding. Medium 3.5 is the company’s “first flagship merged model,” and it replaces Mistral Medium 3.1 and Magistral in Le Chat, while also replacing Devstral 2 in Vibe. Basically, Mistral wants one strong general checkpoint that can do quick chat, lon(mistral.ai)alist every time. (mistral.ai) ### How good is it on the tasks Mistral cares about? Mistral is leaning hard on coding and agent workflows. The company says Medium 3.5 scores 77.6% on SWE-Bench Verified and 91.4 on τ³-Telecom, and says it beats Devstral 2 across its coding benchmarks. Those are vendor-picked benchmarks, so they are not the whole story, but they do show the intended target: s(mistral.ai 1)(mistral.ai 2) ### Why is the four-GPU claim the real headline? Because that is the part that changes who can use it. “Open weights” matters less if the model still needs a giant cluster to be useful. Mistral is saying this one can be self-hosted on as few as four GPUs, which pushes it into reach for more startups, internal enterprise teams, and regional clouds. Think of it (mistral.ai)ture choice.” (mistral.ai) ### What’s the catch? The license is not plain Apache 2.0. Mistral says Medium 3.5 uses a modified MIT license with exceptions for companies above a large-revenue threshold. So this is open-weight and commercially usable for many developers, but not frictionless in the way a fully permissive Apache release is. That is a meaningful difference if you are a hyperscaler or a very large enterprise. (huggingface.co) ### Why does this matter beyond Mistral? It sharpens Mistral’s strategy. Earlier this year, Mistral’s flagship open release was Large 3 — a sparse MoE model under Apache 2.0, optimized for bigger nodes. Medium 3.5 takes a different angle: less about maximum scale, more about collapsing capabilities into a model that is easier to deploy broadly. That gives the o(huggingface.co).” (mistral.ai) ### Bottom line Mistral did not just launch another model. It launched a claim — that a 128B dense system can be open, useful, and runnable enough to compete for real workloads. If that holds up in practice, more of the AI stack moves off locked-down APIs and onto infrastructure people can actually control. (mistral.ai)
