Mistral's "Le Chonk" Proves Open AI Is Missing One Thing

Mistral just announced Mistral Large 4, nicknamed “Le Chonk”: about a trillion parameters, available through Mistral’s API now, with the downloadable model files (open weights) promised for the end of October.

Open weights matter. You could run a frontier-class model on your own servers and keep your data in the building.

But here’s what the headline skips. In Matthew Berman’s testing, the model itself wasn’t the problem. Connecting it to agent tools was.

Closed AI apps ship with the plumbing installed: tool connections, memory, an app to run it all. Open models hand you the engine, and you build the car. That missing layer, the “harness,” may be where the real opportunity sits.

The test worth running before you switch:

1. Pick one small job (for example: meeting note → task list)
2. Run it through an open model
3. Run the same job through a managed agent platform
4. Compare setup time, failures, privacy, and cost

Not benchmark scores. Your workflow.

Fine print: the open-weight release date and license haven’t been fully confirmed yet, and benchmark claims are Mistral’s own.

Which would you trust with your workflow: open or managed?

Mistral’s Large 4 announcement (October 6, 2026): Mistral AI on X and mistral.ai/news Matthew Berman’s Mistral Large 4 test video (YouTube, October 7, 2026): [add link] Launch coverage: https://pasqualepillitteri.it/en/news/21196/mistral-large-4-launches-open-weights-late-october