Mistral Large 3 Goes Open Source Under Apache 2.0
Mistral AI has released the weights for Mistral Large 3, a 70-billion-parameter model, under the permissive Apache 2.0 license. The model is available on Hugging Face and reportedly outperforms Llama 4 on several multilingual benchmarks.
Original sourceMistral AI has open-sourced Mistral Large 3, releasing the full model weights for its 70-billion-parameter flagship under the Apache 2.0 license — one of the most permissive licenses available for a model of this size and capability. The weights are downloadable directly from Hugging Face, allowing developers, researchers, and companies to run, fine-tune, and deploy the model commercially without royalty obligations or usage restrictions tied to Mistral's platform.
The release positions Mistral Large 3 as a direct challenger to Meta's Llama 4 in the open-weight model space. According to Mistral, the model outperforms Llama 4 on several multilingual benchmarks, though the specific benchmark methodology and datasets have not yet been independently verified. The 70B parameter count places it in the range of models that can realistically be run on high-end consumer or prosumer hardware with quantization, expanding the practical deployment surface beyond data center infrastructure.
For enterprise teams evaluating open models, the Apache 2.0 license removes a common legal friction point that more restrictive licenses like Llama's community license introduce. This matters particularly for companies building products where the model's license terms interact with their own IP policies. Mistral has consistently used permissive licensing as a strategic differentiator against both closed API providers and Meta's more conditionally licensed releases.
Mistral Large 3 follows the company's pattern of releasing commercial API models and subsequently open-sourcing them — a cadence that keeps Mistral relevant in both the commercial and open-source communities simultaneously. Whether the multilingual benchmark claims hold up under independent evaluation will determine how seriously this model is adopted relative to alternatives like Llama 4, Qwen 2.5 72B, and DeepSeek V3, all of which occupy similar capability and deployment profiles.
Panel Takes
The Builder
Developer Perspective
“Apache 2.0 on a 70B model is the correct default — no usage caps, no 'community license' gotchas, no legal review required before you ship. The primitive here is clean: download weights, run inference, fine-tune on your data, deploy in your product. What I want to see before getting fully excited is actual quantized performance numbers at Q4 and Q8 — a 70B model is still a serious infrastructure commitment and the gap between benchmark conditions and real-world serving latency is where these releases usually lie.”
The Skeptic
Reality Check
“'Outperforms Llama 4 on several multilingual benchmarks' is doing a lot of work in that sentence — Mistral designed the evaluation, and 'several' is not a methodology. The Apache 2.0 license is genuinely valuable and I won't dismiss that, but in a field where Qwen 2.5 72B and DeepSeek V3 are already available at similar parameter counts with strong independent evaluations, Mistral needs third-party reproducible benchmarks, not a press release. I'll revisit when someone not named Mistral runs the evals.”
The Futurist
Big Picture
“The real bet Mistral is making here is that the Apache 2.0 license itself becomes a selection criterion as enterprise legal teams increasingly audit model supply chains — and that bet is already paying off. The second-order effect worth watching: when a 70B Apache-licensed model is competitive with frontier closed APIs, the 'just call the API' default starts to break down for any company that cares about data residency, cost at scale, or IP ownership. Mistral is riding the trend of model capability commoditization on purpose, positioning the license as the moat when the weights become table stakes.”
The Founder
Business & Market
“Mistral's cadence of commercial-then-open-source is a legitimate strategy — they monetize the early adopters on the API, then open-source to build ecosystem gravity and block competitors from owning the open-weight segment. The Apache 2.0 license is the product here as much as the weights, because it lets enterprises use this without a procurement conversation, which accelerates adoption in exactly the accounts that eventually buy Mistral's enterprise offerings. The risk is that this cadence only works if each new commercial release stays meaningfully ahead of what they just open-sourced — and that window is compressing fast.”