Le Chonk Preview Runs Ahead of Open Weights

Mistral Large 4, branded Le Chonk, is live in public API preview as a 1.05-trillion-parameter multimodal mixture-of-experts model while open weights remain on a late-month clock. Teams that want agentic and cyber-leaning workloads can call it now without waiting to self-host.

The public preview opened on 6 October 2026. Open weights are still scheduled for the end of October 2026, so the hosted path is the only production route until that drop.

6 October through the API-first sequence

On 6 October 2026, Mistral AI put Mistral Large 4 into public preview on its API and framed Le Chonk as a 1.05-trillion-parameter multimodal MoE. The company pitch stressed agentic use and cyber-facing evals, including a Cybench score of 93% cited in launch materials, plus a Finance Agent v2 showing on the model card narrative. Same-day write-ups from Decrypt, MarkTechPost, and CNET covered the launch.

12 October brought a second MistralAI launch thread that pushed the video and model framing back into the feed. The original launch video had already drawn 5.6M+ views on X. That engagement explains why the preview read as a time-sensitive story rather than a quiet model-id swap.

Weights are still promised for the end of October 2026. Until that drop, the only production path is the hosted API. The chronology runs in three beats: preview open, press and social amplification, open weights still pending.

End of October remains the weights gate

Access is the half of the sequence that already shipped. Public API preview means developers and companies with Mistral accounts can route traffic now. It does not mean unrestricted open weights, local fine-tunes, or air-gapped deploy. If your constraint is data residency or offline inference, the October 6 ship did not solve it. If your constraint is getting a strong multimodal MoE into an agent stack this week, the API path is the one that opened.

Buyer fit is simple at this stage of the timeline. Production teams with budget and latency tolerance for a hosted endpoint can start. Labs that only move when weights and licenses clear should hold. Teams that need self-host or offline inference should wait for the end-of-October weights drop rather than treat the preview as a permanent deploy shape.

1.05 trillion parameters still behind an API wall

Read as a pattern, the October 6 preview plus the late-October weights promise looks like a deliberate two-step: capture API revenue and evaluation mindshare while the open release is still on the calendar. Shipping Le Chonk closed-first creates a short monetization window before the open drop lands.

What's underappreciated here: the preview is the live product for anyone who cannot wait until weights land, and the weights date is the competitive pressure valve. If the end-of-October open release holds, Mistral gets both a paid runway and a self-host story. If it slips, the API path starts looking like a tax on patience.

Third-party coverage already framed Large 4 around agent and cyber strength rather than chat novelty. That framing will stick only if independent leaderboards and customer workloads confirm the Cybench-led pitch after the launch week noise fades. We do not yet have a full independent scorecard dump that replaces the company’s own cited figures. Treat 93% Cybench as a sourced launch claim, not as a settled field ranking.

The next move implied by the sequence is operational, not theatrical. Expect more agent-template marketing, more price-table comparisons against US frontier APIs, and a hard calendar watch on the weights drop. The pattern breaks if that watch turns into vapor.

1 November without published weights

The read above fails on a concrete clock. If 1 November 2026 arrives with no published Large 4 weights, no license text, and no firm reschedule from Mistral, the October sequence collapses into a hosted product launch. Preview buyers would then be paying for Le Chonk as an API-only model, with self-host still unavailable on the stated schedule.