Wait Which Model?
← Back to directory
Mistral AI

Mistral Large 4

Released Oct 6, 2026 · knowledge cutoff unpublished

Status
Unknown
Location
France
Type
LLM
Modality
Multimodal
Context window
1M
Max output
—
Speed
116 tok/s · 19s to first answer token
Price ($/MTok in / out)
$1.36 / $4.18
Cost per task
$1.13
Open weights
No
Benchmarks0 of 10 reported

MMLU-Proretired

—

GPQA Diamond

—

SWE-bench Verified

—

SWE-bench Pro

—

Terminal-Bench 2.1

—

AIME (latest)retired

—

Humanity's Last Exam

—

LMArena Elo

—

GDPval-AA v2

—

ARC-AGI-2

—
Strengths & weaknesses

Strengths

  • Trained across more than 160 languages including every official EU language, so it is built for multilingual European deployments rather than English-first work
  • Strong at visual grounding — it locates and verifies objects in dense scenes, technical drawings, PDFs and very large satellite images rather than only describing them
  • Takes on vulnerability reproduction and patching work that refusal-prone closed models decline — useful for defensive security teams

Weaknesses

  • A public preview whose reinforcement-learning run is still in flight — Mistral expects large improvements in the coming weeks, so behaviour is a moving target, not a fixed release
  • Reasons before every answer, so replies take a long time to start even on simple prompts
  • Weights are promised for the end of October but not yet out, so self-hosting has to wait

Public preview announced 2026-10-06 on Mistral Studio; Mistral says the weights will be released by the end of October and names no licence yet, so openWeights is false and license null until they ship. Mistral's two official pages disagree on size: the announcement says 1T total / 49B active parameters, while the docs model page says 1.05T total / 52B active plus a 1.6B vision encoder. Context is 1M per Mistral's docs (Artificial Analysis lists 524,288). Pricing is the $1.36 / $4.18 list rate ($0.14 cached input); the docs page currently shows exactly half ($0.68 / $2.09 / $0.07 cached) with the list price struck through, and gives no end date for the discount. Mistral reported none of the tracked benchmark keys: its launch figures are DeepSWE v1.1 61.7%, SWE-Atlas-QnA 59.4%, Terminal-Bench 4 28.3% (not the 2.1 this site tracks), Coding Agent Index 49.8%, AutomationBench 59.9% and Cybench 93%. hle is null: Mistral reports no HLE figure, and Artificial Analysis' 35% comes from its own text-only subset with an unstated tool setting. AA also lists GDPval-AA v2.1 at 1424, which is not the v2 this site tracks. AA measures a single reasoning configuration with no effort levels named, so costPerTask and speed carry no effort.

For developers

API model strings

  • mistralmistral-large-4
  • mistralmistral-large-4-0

Licence

Proprietary — weights not released

Retirement

No retirement announced

Lineage

No recorded predecessor

News