← Back to directory
Mistral AI
Mistral Large 3
Released Dec 1, 2025 · knowledge cutoff unpublished
- Status
- Unknown
- Location
- France
- Modality
- Text
- Context window
- 262K
- Max output
- 66K
- Speed
- —
- Price ($/MTok in / out)
- $0.50 / $1.50
- Cost per task
- $0.060
- Open weights
- Yes
Benchmarks1 of 8 reported
GPQA Diamond
67.2%
MMLU-Pro
—
SWE-bench Verified
—
Terminal-Bench 2.1
—
AIME (latest)
—
Humanity's Last Exam
—
LMArena Elo
—
ARC-AGI-2
—
Strengths & weaknesses
Strengths
- A fast, broad-coverage generalist tuned for throughput rather than depth
- Good backbone for multilingual retrieval over long documents and messy PDFs
- Apache-licensed, so it can be self-hosted anywhere without negotiation
Weaknesses
- Pattern-matches rather than reasons — it doesn't improve if you let it think longer
- Context feels short for the long-document work it is otherwise best at
Official pricing $0.50/$1.50; some trackers list hosted rates of $2/$6. GPQA Diamond 67.17 per Mistral's HF model card; not a dedicated reasoning model, so most 2026 reasoning benchmarks are unreported.
For developers
API model strings
Not researched
Licence
Not researched
Retirement
No retirement announced
Lineage
No recorded predecessor