Mistral's Large 4 technical report: agent benchmarks and a cache trade-off
From the AI digest for Nov 3, 2026
What happened: Mistral's own posts announce the Large 4 technical report, published on alphaXiv. Mistral's Pierre Stock adds that Large 4 ("Le Chonk") is "now open-weight" and, by Mistral's framing, far more efficient and compact than its Preview three weeks earlier. A same-day follow-up post says Large 4 now sits in the lineup on Open Math OS — "1.05T open weights, 1M context, live today for approved teams" — Mistral's open-models research workspace, with access gated to approved teams. All vendor claims — no independent coverage yet.
What's genuinely new: The report's central technical finding: aggressively cutting KV-cache memory by 16× hurt long-context agentic tasks — where a model must retrieve information from much earlier context — even when language-modeling scores looked unchanged. Mistral's fix was to restore full-context attention to more layers and adopt DeepSeek Sparse Attention to select the 2,048 most relevant tokens. Vendor-run scores for the resulting 1T-parameter MoE model (52B active): 41% on Terminal-Bench 4.0 and 19.7% on FrontierCode 2.0, with the NVFP4-quantized checkpoint said to fit on four GB300 GPUs.
Watch: independent reruns of the agentic scores, confirmation of the open-weight release and Open Math OS's approval gate and licensing terms beyond the vendor's posts, independent checks of the 1M-token context claim, and whether the cache-agency trade-off finding gets picked up by other labs.
Links: Mistral AI · Mistral AI · Mistral AI · Nov 3, 12:30 PM ET