Mistral Large

Mistral Large – Flagship Single-Node Dense Model

Mistral Large is Mistral AI's top-tier commercial model line, positioned for high-complexity reasoning, multilingual tasks, and long-context applications. Its defining engineering constraint is single-node inference suitability: at 123B parameters (Large 2), it is explicitly sized to run at large throughput on a single node — a deliberate departure from the multi-node requirements typical of frontier-scale models.

Architecture & Training Focus

Large 2 was trained on a heavily code-weighted corpus following Mistral's experience with Codestral. A specific training objective was response conciseness — Mistral officially documented and benchmarked output length against Claude 3 Opus, Claude 3.5 Sonnet, Llama 3.1, and GPT-4o, targeting shorter, production-suitable responses.

Version Lineage (Official)

• Large (2402): First release; 32k context, native function calling, Azure launch partner
• Large 2 (2407): 123B, 128k context, code-heavy training, hallucination-reduction focus
• Large 2.1 (2411): Top-tier high-complexity release, November 2024
• Large 3 (2512): Sparse MoE, 41B active / 675B total parameters, Apache 2.0