Breaking

Mistral Large 4 Enters Agent Arena as Mistral Opens Public Preview of Its 1 Trillion-Parameter Model

October 6, 2026 at 10:22 EDT

Foundation Models
AI Agents
Open Source

Mistral Large 4, Mistral AI's new open-weight flagship model, has been added to Agent Arena for community testing while Mistral runs a public preview through its API. Arena said votes will shape the model's agentic evaluation, with scores still to come.

Oct 6, 2026 · Mistral Large 4

A 1T open-weight model enters agent testing

Mistral Large 4 is being evaluated for tool-use performance. The story now moves from launch benchmarks to long-horizon agent results.

1.05T

total parameters

49B

active parameters

token context

2.16M

agent sessions on the board

Total capacity vs active compute

Using roughly 50B parameters per block, the active path is a small slice of the full MoE model.

1.05T

total

49B

active

MoE tradeoff: a very large parameter pool is available, while each request activates only part of it.

Agent task

Complex workflow

→

Tool use

Search, files, terminal

Outcome

Completion and steerability

Why users care

Open weights, European infrastructure and strong claims in coding, cyber and enterprise workflows make ML4 a practical deployment candidate.

What remains open

Independent results and long-horizon agent scores will decide whether launch benchmarks hold up in practice.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.