Mistral 7/12 Mistral Large 3: 145B dense flagship with 128K context and Apache-2.0 weight drop

2 sources

Mistral on 2026-07-12 introduces Mistral Large 3 — a 145B-parameter dense flagship at 128K context with native function-calling + JSON-schema tool-use and a chat-instruct variant. Self-reported figures: MMLU-Pro 79.2, GPQA-Diamond 71.6, HumanEval+ 93.4, MultiPL-E 88.1, MT-Bench v3 9.21, IFEval strict 88.4. License is Apache-2.0 with a Mistral-research addendum; deployment via Mistral La Plateforme + Azure AI Foundry + Bedrock.

Why now

It lands on the same day as Qwen3-Max (China) and Kimi K2.7-0913 (Moonshot, China) — three vendor-aligned model drops in a single 24-hour window, with Mistral as the European third. Creators can carry the trio as one beat instead of three.

Why it is worth publishing

Best for European-data-sovereignty creators + dense-vs-MoE comparisons. The Mistral-research addendum is the legal hook; the BF16-vs-MXFP4 weight-format detail is the deployment hook.

Evidence basis

Single official event + dense-flag positioning + Apache-2.0 + European-vendor trust positioning. Medium heat, niche-but-clear audience.

“On the same day Alibaba shipped a 1.2 trillion-parameter MoE, Mistral shipped a 145 billion-parameter dense model — both Apache-2.0, both with tool-use, both a single 24-hour window apart.”

Angle

Frame Mistral Large 3 as Mistral's first Apache-2.0 dense flagship at the 145B scale with native tool-use — position the dense-vs-MoE contrast against the same-day Qwen3-Max (1.2T MoE) instead of against everything released in July.

Format

Carousel comparison + a single-card distillation

Demo idea

Two-card comparison: row 1 — 'Alibaba Qwen3-Max: 1.2T total / 32B activated MoE, 1M context'; row 2 — 'Mistral Large 3: 145B dense, 128K context'; columns: total params / activated params / context / type / license / addendum. Use this as a hook for the dense-vs-MoE comparison that you can re-run on any future mixed-vendor release day.

Platform notes

Self-reported + addendum (medium risk): the Mistral-research addendum boundary was not extracted in the captured summary — quote only what the Mistral news post and the HF card state. When viewers ask 'is Mistral the European answer to US frontier labs,' don't generalize from a single release.

Usable claims

Mistral Large 3 is Mistral's 145B-parameter dense flagship with 128K context, native function-calling plus JSON-schema tool-use, and a Mistral-medium-style chat variant; self-reported figures include MMLU-Pro 79.2, GPQA-Diamond 71.6, HumanEval+ 93.4, MultiPL-E 88.1, MT-Bench v3 9.21, and IFEval strict 88.4; weight formats include BF16 sharded safetensors and MXFP4 inference snapshots; license is Apache-2.0 with a Mistral-research addendum; deployment is via Mistral La Plateforme + Azure AI Foundry + Bedrock.

Evidence pipeline

From the news

Mistral 7/12 Mistral Large 3: 145B dense flagship at 128K context with native tool-use

Breakdown

Walks Mistral Large 3's spec (145B dense, 128K context, MMLU-Pro 79.2 / HumanEval+ 93.4 / MT-Bench v3 9.21 / IFEval strict 88.4) paired with the BF16-vs-MXFP4 weight-format split and the Mistral-research addendum boundary — without stating the addendum's scope, which was not extracted. Read alongside the same-day Qwen3-Max explainer for the dense-vs-MoE contrast.

Sources

Risks

Pin the link to each vendor page and the companion HF model card; quote only what those pages state; do not paraphrase self-reported benchmark numbers as third-party validation; for Kimi K2.7-0913 specifically, frame as a re-tag of existing weights, not a base-model swap.
Pin the link to each license text; quote only what the page states; do not paraphrase Apache-2.0 as the sole operating condition; flag addenda by name without stating their scope.

Demo ideas

Pair Mistral Large 3 with Qwen3-Max (same day, both Apache-2.0) as 'open-weight governance on 7/12' — flag the addenda boundaries carefully and avoid equating 'Apache-2.0' with 'fully unrestricted'.
For long-form European / data-sovereignty coverage, build a 'Mistral — Qwen — Moonshot trio' triad page where Mistral sits as the European anchor and Qwen + Kimi as the Chinese anchors, with one row per model and one addendum column.
If you do MXFP4 vs BF16 coverage, anchor it on the Mistral Large 3 weight-format split — the HF card lists BF16 sharded safetensors + MXFP4 inference snapshots; quote only what the card states, run your own quality comparison if you generalize.