Mistral Large 3
Mistral AIMULTIMODAL Open weight
- Provider
- Mistral AI
- Modality
- MULTIMODAL
- Parameters
- 675B
- Context window
- 256,000 tokens
- Weights
- Open
- Released
- 2 Dec 2025
Mistral Large 3 (Dec 2025) is Mistral's Apache-2.0 frontier flagship — a 675B/41B-active multimodal MoE with strong multilingual reach at a fraction of US-frontier API cost.
Best for
- Broadest general-purpose multimodal + multilingual capability
- Best image understanding in the Mistral line
- Now Apache 2.0 — full commercial self-host rights
- Top-end answer quality when compute budget allows
How it compares — and the India angle
- At $0.50/$1.50 it undercuts GPT-5-class and Claude Sonnet output pricing by ~5-10x.
- Officially supports 9 Indic languages (Hindi, Bengali, Gujarati, Kannada, Marathi, Nepali, Punjabi, Tamil, Telugu).
- Apache 2.0, but 675B is too large for cheap local GPUs — practical mainly via API; no rupee billing.
Benchmarks
Representative public scores (approximate, higher is better) for relative comparison. Check the provider for the latest official results.
GPQA Diamond67
How to access
API ~$0.50 / 1M input, ~$1.50 / 1M output. Apache 2.0 (free to self-host; needs a serious GPU cluster).
Access Mistral Large 3