AINEWS
Mistral's open weight (downloadable trained model weights, not open training code) Large 4 runs 49B active parameters out of 1.
Mistral Large 4 entered public preview on October 6, 2026, as an open-weight multimodal model with a one-million-token context window and a Standard API price below a dollar per million input tokens (Mistral model page, Mistral announcement). The release expands the supply of frontier-class open-weight options at sub-dollar API rates, which changes the build-versus-buy math for teams that could not justify closed-vendor pricing at scale.
The model is a sparse mixture-of-experts with 49 billion active parameters out of 1.05 trillion total, paired with a 1.6-billion-parameter vision encoder (Mistral model page). In MoE terms, only 49B parameters fire per token even though the model holds roughly a trillion in storage. That is the engineering tradeoff the architecture reveals: a builder pays the memory cost of 1.05T parameters for the compute cost of 49B. Mistral trained the model from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacenters, the company says (Mistral model page).
The price sits in plain terms. The Standard tier lists $0.68 per million input tokens, $0.07 per million cached-input tokens, and $2.09 per million output tokens, with crossed-out legacy rates of $1.36, $0.14, and $4.18 next to the new numbers (Mistral model page). The cheapest closed-vendor flagships charge several dollars per million input tokens, the comparison the model page implicitly invites. Third-party analysis frames Large 4's effective cost-per-useful-token against closed-vendor flagships and open-weight peers as the story for builders deciding on inference budgets (Kingy.ai coverage). Preview pricing can shift at general availability, and Mistral has not committed to keeping these rates.
Independent benchmark snapshots are useful, with the usual caveat that vendor-reported scores are marketing until reproduced. On DeepSWE v1.1, a software-engineering task suite, Mistral reports 61.7 percent, behind Kimi K3 and DeepSeek V4.1 Flash on the same test (CellCog analysis). On the Vals AI broad industry index, Large 4 lands at 48.05 percent, behind GPT-6 Astra, GPT-6.1 Sol, Claude Sonnet 5.5, and Claude Opus 5.5 (Kingy.ai coverage). Mistral's own claim that it "significantly outperforms any open-weight model developed in the US or Europe" is the marketing line; whether it holds against Qwen, DeepSeek, and the Llama lineage is the open question for independent evaluation (Mistral announcement).
The license is the part that decides what a builder can actually do. The model page currently lists the license as "Open," with final terms deferred until the weights land at the end of October 2026 (Mistral model page). Mistral's prior open-weight model, Large 3, shipped under Apache 2.0, and most of Mistral's open-source models follow that pattern, with a small number under a modified MIT license that adds a commercial-use trigger at $20 million monthly revenue (Mistral help center, IntuitionLabs license analysis). The Mistral Nonproduction License, or MNPL, blocks any commercial activity without a separate Mistral agreement regardless of users or revenue. The "open-weight" answer therefore depends on which bucket the late-October release lands in. A permissive Apache 2.0 would put a frontier-class multimodal model into the same self-hosting lane as the rest of the open-weight stack. A modified MIT or MNPL carve-out would change the calculus for any team crossing the revenue threshold.
The community read so far is guarded. Hacker News discussion on the launch thread picked apart the marketing claim and asked for head-to-head reproductions against DeepSeek and Qwen on coding and reasoning tasks, the gap any independent eval would have to close (Hacker News thread). The preview is live on Mistral Studio now. The weights drop at the end of October. The license answer is the next thing that turns this from a product page into a deployment decision.