$ cat wiki/models/mistral-large-3.md
Mistral Large 3
modelupdated 2026-07-21created 2026-07-12
Spec
| Attribute | Value |
|---|---|
| Developer | Mistral AI |
| Released | not yet (early access since 2026-07-06; broader GA expected mid-to-late July 2026) |
| Announced | 2026-07-04 (CEO confirmation in TechCrunch profile) |
| Context window | unknown |
| Pricing | unknown |
| License | Apache 2.0 |
| Availability | Early access (July 6, 2026); broader release expected mid-to-late July 2026 |
| Status (as of 2026-07-12) | Early access (July 6, 2026); broader GA expected mid-to-late July 2026 |
| Architecture | Mixture of Experts (MoE) — "fat but sparse" |
| Total parameters | ~675 billion |
| Active parameters per token | ~41 billion |
| Benchmarks | Not yet published (pending GA) |
Release Date
Early access opened July 6, 2026. CEO Arthur Mensch confirmed the model on July 4, 2026 in a TechCrunch profile. Broader release expected mid-to-late July 2026 (no firm date as of July 12).
(source) (TechCrunch)
Timeline
- July 4, 2026: CEO Arthur Mensch confirms the model as "fat but sparse" MoE; ARR $400M+
- July 6, 2026: Early access opens (developer sign-up)
- Mid-to-late July 2026: Broader GA expected
Architecture
Mensch described the design as "fat but sparse": a large total-parameter MoE with sparse per-token activation. This approach gives the model:
- High capacity: 675B total parameters → large knowledge base and reasoning depth
- Low inference cost: only ~41B parameters active per forward pass → cost-efficient serving
- Comparable to: DeepSeek-V3 (671B/37B active), GPT-MoE-class models
(source)
License
Apache 2.0 — Mistral's most permissive frontier-class open-weight model ever. Enables commercial deployment without usage restrictions.
Benchmarks
Not yet published. Benchmarks expected at GA. Mistral's CEO positioned this as a direct competitor to the closed-frontier tier (GPT-5.6, Fable 5), which would make the benchmarks especially significant.
Positioning
- Direct open-weight competition with DeepSeek-V3 (671B/37B active) and Llama 4 Ultra
- Apache 2.0 license targets European enterprise sovereignty use cases
- Follows Leanstral 1.5 (July 1–2, theorem proving) and Robostral Navigate (July 8, robotics) in an unusually dense July 2026 release cadence
Compared To
| Model | Params (total/active) | License | Status |
|---|---|---|---|
| Mistral Large 3 | 675B / 41B | Apache 2.0 | Early access (July 6) |
| DeepSeek-V3 | 671B / 37B | MIT | GA |
| Leanstral 1.5 | 119B / 6.5B | Apache 2.0 | GA |
| Devstral 2 | 123B / ? | Mistral Research | GA |
Related
- Mistral AI — developer
- Leanstral 1.5 — sibling open-weight model (July 2026, formal verification)
- Devstral 2 — coding-focused open-weight model
- Reasoning Models — relevant if Deep Think-style reasoning is confirmed