Mistral Large 4
Where it stands: Mistral's Large 4, nicknamed Le Chonk, is a 1-trillion-parameter open-weights model with a 1M-token context that Mistral says is natively multimodal. Its first third-party score of 38 sat one point behind Muse Spark Open, but a revised Artificial Analysis index weighted toward private tests cut it to 25, below DeepSeek V4.1 Flash and GLM-5.3. Mistral's own technical report claims 41% on Terminal-Bench 4.0 and 19.7% on FrontierCode 2.0, and the model is live on Open Math OS for approved teams. Watch for independent reruns and confirmed weights and licensing.
Latest
- 2026-11-03(Mistral's own posts — x.com/MistralAI + x.com/MistralAI): the Large 4 technical report is out on alphaXiv, with a KV-cache finding — cutting cache memory 16× hurt long-context agentic performance even with unchanged language scores; Mistral's fix was restoring full-context attention to more layers plus DeepSeek Sparse Attention selecting the 2,048 most relevant tokens — and vendor-run scores for the 1T-parameter MoE (52B active): 41% on Terminal-Bench 4.0, 19.7% on FrontierCode 2.0; the NVFP4 checkpoint fits four GB300 GPUs. Pierre Stock: "ML4 is now open-weight," "way more efficient and compact" than the Preview three weeks earlier. Vendor claims throughout. Watch: independent reruns of the agentic scores, weights and licensing confirmation, and whether the cache-agency trade-off finding is picked up elsewhere. Development of the Oct 6/16/19/28 entries in this storyline.
AI image - 2026-11-03 (Mistral — x.com/MistralAI): same-day follow-on from the technical-report posts — Large 4 joins the model lineup on Open Math OS: "1.05T open weights, 1M context, live today for approved teams." First concrete distribution detail beyond the report: availability runs through Mistral's open-models research workspace and is gated to approved teams; the tweet's 1.05T matches the report's 1T-class MoE (52B active). Vendor-stated throughout. Watch: independent confirmation of the 1M-token context, Open Math OS's approval criteria, and weights/licensing terms.
- 2026-10-28 (~6:00 PM ET, Actu IA — actuia.com/actualite): published coverage of the v5 revision quantifies the reset — twelve already-scored models lose 12–16 points under an index now weighted 60% toward private tests; Mistral Large 4 ("Le Chonk") falls from 38 to 25, dropping the Oct 6 open-weights flagship below the 32s of DeepSeek V4.1 Flash (max) and GLM-5.3 (max) on the same board. Both of Le Chonk's earlier numbers in the thread (AA 38 on Oct 19, one point behind Muse Spark Open; the Oct 16 STEMBench claim) are pre-revision scores now superseded. Watch: any published v5 table naming the other eleven casualties, independent commentary on the 60%-private weighting, and whether any released-weights model regains ground in a later v5 revision.
October 2026 · 3 updates
- 2026-10-19 (~4:00 PM ET, Actu IA — actuia.com/actualite): the thread's standing "third-party evals pending" watch on Mistral Large 4 gets its first published result — Artificial Analysis puts Muse Spark Open at 39, one point ahead of Large 4 at 38, while Large 4 wins the agent tasks; the capture truncates an additional Large 4 "1.8" figure (likely a speed/writing metric — unconfirmed). Same-day adjacent, same band: Tencent's Hy-Memory-30B-A3B, a free open-source mixture-of-experts model, claims Claude Opus 5 parity on AI agent memory — on its own benchmarks (~2:00 PM ET, MSN — msn.com/en-us); the vendor-benchmark genre's self-graded caveat applies verbatim. Watch: the full AA table and field context, the truncated Large 4 figure, independent runs of Hy-Memory's parity claim, and release of its eval scripts.
- 2026-10-16 (~7:29 AM ET — x.com/MistralAI): Mistral's own channel on Mistral Large 4 / "Le Chonk" — "Turns out le Chonk is into rocket science. Mistral Large 4 scoring well in engineering and astronomy on Ulam's private benchmark, STEMBench" (stembench link carried in-post). Vendor claim on a private benchmark, no scores printed in the post; a development of the Oct 6 open-weights/Surge-AI-eval entry in this thread. Watch: any published STEMBench table, independent reruns, and whether the "private benchmark" framing draws the same benchmark-integrity skepticism the thread has logged elsewhere.
- 2026-10-06 Mistral Large 4, "Le Chonk" (Mistral announcement + The Register): vendor-claimed 1T parameters, natively multimodal, open weights — the open-weights race's European frontier-class entrant; blind Surge AI human eval: #1 among open-weights, #2 overall behind Opus 5; Gostev (via LeCun): trained on 4,000 GPUs vs Astra's 100,000 — the compute-disparity frame doing anti-scaling work. Vendor-claimed numbers; independent runs pending. Watch: weights' licensing/ spread, third-party evals, whether the blind-human-eval announcement move spreads.