Mistral opens a trillion parameters, with one string attached

The big story

Mistral Large 4 ("Le Chonk") is a natively multimodal MoE with about 1.05T total parameters, 49B active per token (52B counting embeddings and output head), 128 routed experts with 4 active plus one shared, MLA with DeepSeek-style sparse attention, and 1M context. It ships as FP8 in 54 shards (about 1.06 TB) plus an NVFP4 checkpoint; ggml-org and Unsloth had GGUFs up the same day. Mistral's own numbers: 48% on Artificial Analysis's Coding Agent Index, tied with GLM-5.3, using 47% fewer tokens per task, and first among open models on Cognition's FrontierCode. On AA's overall intelligence index it sits fifth among open-weight models, behind the Chinese labs — Mistral says so itself. This is not the October 6 preview checkpoint; three more weeks of RL went in and Mistral claims gains had not saturated.

The license is the part this audience should read twice. Large 3 shipped Apache-2.0 in December 2025. Large 4 ships under the "Mistral Ecosystem Open-Weight License," which permits use, fine-tuning and commercial products but requires anyone who "operates a Model as a Service business" to sign a separate commercial license first. That is a Llama-style carve-out wearing MIT clothing. For self-hosters nothing changes; for the inference providers who would otherwise make a 1T model cheap to try, it does, and it means the weights-available-but-restricted column of the ledger just gained its biggest entry. The "sovereign European model" pitch also landed the same week POLITICO reported a German search engine dropping Mistral for Chinese open models. The buyers Mistral wants are the ones most sensitive to MIT-versus-not.

New open-weight releases

• Mistral Large 4 (~1T MoE, 52B active; Mistral Ecosystem Open-Weight License) — above. Card.
• Ling-3.1-flash-1019 (560B MoE, 25B active; MIT) — Ant's refresh: 574 GB of FP8 in 54 shards, plus a 1.14 TB BF16 base model, which almost nobody ships at this scale. Hybrid-linear architecture (63 KDA layers, 9 gated-MLA), 512 experts. The card says 1M context; the config says 256K. AA index 44, up from 41. Card.
• Nemotron 3.5 Transcribe 30B-A4B (34B actual, 4.3B active; OpenMDW-1.1) — NVIDIA's Mamba-2/Transformer hybrid MoE speech LLM: 40 languages, hour-long recordings, 64K context, speaker attribution and translation in one checkpoint. 4.71% mean WER across nine English sets, top open model on the HF Open ASR leaderboard; the Granary v2 training set (463K hours, 48 languages) came with it. A community prune already took it from 67.5 GB to 10 GB at 5.16% WER. Card.
• LFM3 (1.7B text; 2B VL, Audio, Omni; two decision models; LFM license) — Liquid's third generation, six models from one 1.73B base trained on 22T tokens, 256K context, tri-hybrid (gated short conv, sliding-window and global attention). Claims to beat Qwen3.5-9B on agentic tasks at ~1 GB in 4-bit. Zero-day llama.cpp support including the omni model, and a video-input PR already merged. The LFM license is free only below a revenue threshold — not open source. Blog.
• Nanbeige4.3-33B (33B dense) — claims to beat Qwen3.8-27B across the board; the dense option in a MoE month, with a quant ladder down to 13.7 GB. License not prominent — check the card.
• Clover-35B-A3B-Preview (35B MoE, 3B active) — DiffuSpace's first release, a diffusion LLM claiming state of the art among open diffusion LMs. Card.
• Xing4.0-29B-A4B (29.5B MoE, ~4B active) — China Telecom's TeleAI, MLA plus 4-stream mHC residual mixing; an ik_llama.cpp PR gives it a native runtime.
• OCR, all at once — Baidu Unlimited-OCR-2 (sparse-reference sliding window for unbounded pages), OpenBMB MiniCPM-OCR-1.0 (1.3B, 93.68 on real-world parsing, plus a judge model for OCR errors), Xiaomi-OCR-0 (0.8B), and Surya 2.5 (689M, 85.9 OmniParseBench).
• Also: Tencent Hy-SimulMT-30B-A3B (simultaneous translation, GGUF day one), NVIDIA Nemotron 3.5 Bolt E4B (Gemma 4 E4B post-trained), ServiceNow Apriel-Voice-27B-preview, SenseNova-Loop-9B-Preview.

Policy & politics

Cohere said on October 26 that Qwen3.8-Flash-Next, during a CTI-REALM run, tried to upload a test string to an Alibaba Cloud bucket in Beijing after stating in its reasoning it would use no tools. The sandbox blocked it; Cohere says it happens under once per 100,000 tool calls and that it cannot tell deliberate from training residue. After pushback it released traces and a repro script. LocalLLaMA's read, after going through them: strong evidence the model hallucinates plausible Alibaba URLs, no evidence of a backdoor, and the original claim was prefilled. The lesson is boring and correct — egress-filter your agents — but "Chinese model exfiltrates data" headlines now exist, which matters when CEPA is running pieces titled "The AI Kill Switch" and CNAS is asking whether the door on open weights is closing.

The DOJ is inquiring about Nvidia's $12.9B Hugging Face acquisition, specifically whether it lets Nvidia capture "a fundamental distribution point for AI." The deal is slated to close in the first half of 2027. Whatever the antitrust outcome, the archivist's conclusion is the same one mradermacher reached in the back up your quants thread: the Hub is a single point of failure with a pending change of ownership.

Elsewhere: Alibaba Cloud's CTO reportedly said at Apsara that Qwen4 lands next week, with 1M context and a leaked 4-way hyper-connection residual design; the community thread worth reading is the one worrying about the license, since Qwen3.8-Flash-Next already moved to custom terms. And Artificial Analysis's Index v5 update shows zero open-weight models in its "most attractive" quadrant, which the sub reads as methodology drift; Mistral's own release post leans on the same index.

Ecosystem

Cloudflare disclosed a cross-request KV cache exposure in its GLM-5.3 Workers AI deployment (155 affected responses) — a serving bug, not a model bug, but the first public one of its kind. llama.cpp has a CVE in the PocketTTS listing path; update. On the upside, the fused MoE gate/up CUDA PRs give +10–17% prompt processing on DGX Spark. The DeepSeek-V4.1-Flash-1017 quant war continues: Unsloth's Dynamic v3.1 GGUFs and ISTA-DASLab's 2.25bpw GSQ-RCO each claim the other's comparisons are off, and a 40-day telemetry post shows API providers silently swapped the original V4.1-Flash for -1017 — another argument for local weights. Betelgeuse's expert-prediction trick runs Flash Next at 22–24 tok/s on a 3060 with 16 GB of DDR4, bit-exact. And for pure delight: a RISC-V CPU and then a NES emulator baked into Qwen3.5-4B's weights.

From the index

No new listings this week; two candidates. Ling-3.1-flash-1019 at 574 GB under MIT is exactly the kind of thing we exist to seed, and the base model doubly so. Mistral Large 4 at ~1.06 TB is smaller than Kimi K3 and would be our first modified-MIT Mistral — note that every Mistral in the catalog today (7B, 8x7B, 8x22B) is Apache-2.0. Two existing listings got more interesting: our DeepSeek-V4.1-Flash (MIT, 510 GB, added September 13) is the pre-1017 checkpoint that API providers have since overwritten, so it's now the reference copy of the behavior people are measuring against. And SiliconANGLE reports Hugging Face used GLM-5.2 to fight off an attacker after a commercial frontier model refused — that model is here, under MIT, unlike its 5.3 successor. With the DOJ now circling the Hub, keep seeding.