Models#Deployment#Model inference
Mistral releases Large 4 weights: 1T-param MoE, 606GB after NVFP4
Mistral released the Large 4 weights Oct 29: 1T-param, 52B-active multimodal MoE, top coding scores among open models per its report. NVFP4 checkpoint: 606GB.
Mistral AI released the weights of its flagship Mistral Large 4 on October 29, three weeks after the October 6 public preview. It is one of the largest open-weight models ever published: a natively multimodal mixture-of-experts with 1 trillion total parameters and 52 billion active per token, a 1M-token context window, and, according to the accompanying technical report, the top scores among open models on both coding benchmarks. The release is more compact and efficient than the preview, the company says — but the NVFP4-quantized checkpoint alone weighs 606GB, so “open” has not made it accessible.
Key points
- Architecture: 1T total / 52B active MoE — 4 of 128 routed experts plus 1 shared expert per token, a Pixtral-style vision encoder, native image-text-to-text
- Scores: FrontierCode 2.0 19.7% and Terminal-Bench 4.0 41% (up from 28% in preview) — both coding leads among open models; the Cyber Index 50.9 is near the front but not first, with MiMo-V2.6-Pro at 56% and DeepSeek V4.1 Flash at 57% above it
- Deployment: the official NVFP4 checkpoint is 606GB; the technical report says four GB300 GPUs suffice, and an 8-GPU tensor-parallel node is the default recipe
- License: Mistral Ecosystem Open-Weight License, a modified MIT — commercial use allowed, but selling model-as-a-service needs a separate agreement
- Price: API pricing dropped from the preview’s $1.36/$4.18 to $0.70/$2.10 per million tokens
What the tech report says
The weights are the headline, but the technical report (alphaxiv 2611.mistral-large-4) carries the more interesting engineering story. Mistral’s initial architecture used sliding-window attention, cross-layer attention and YOCO — a design that cuts KV-cache size by roughly 16x. Validation perplexity looked fine, yet SWE-bench-style agentic tasks degraded noticeably: the cache-saving structural choices were directly damaging multi-step tool use. The team performed what it calls model surgery to restore full context, then recovered efficiency with DeepSeek-style sparse attention (2,048 keys retained per query). The report’s stated lesson: evaluate architectures on real agentic workloads, not on language-modeling loss as a proxy. Any team trading memory savings against agent performance will recognize the shape of that mistake.
On the data side, the report describes an agent-built pipeline: task-authoring agents construct executable environments, specs, reference solutions and verifiers across software engineering, cybersecurity and knowledge work; refinement agents then test with multiple solvers, repair broken tests and check for reward hacking, and successful trajectories feed back into training. The cybersecurity strength traces to this system — the Cyber Index score comes from infrastructure that generates and verifies attack-defense tasks automatically.
The cost of actually running it
606GB is the quantized number; the BF16 checkpoint is roughly double that. The report claims calibrated NVFP4 “matches the FP8 reference on evaluations” and fits on four GB300 cards (288GB of HBM3e each). The Hugging Face model card’s default path is less forgiving: vLLM or SGLang with --tensor-parallel-size 8, or torchrun across eight GPUs for native transformers. The card also embeds a DSpark speculative-decoding head (three dense MLA layers) that Mistral says lifts throughput up to 1.9x. Compare Inferact running Kimi K3 on 16 TPU v7 chips from late September: trillion-parameter open models remain an institutional proposition. A four-GPU GB300 node costs millions of dollars; individuals are waiting on community quants. As one Hacker News commenter put it, “I think I need about a terabyte more RAM to run that.”
Against the Chinese open-weight field
This is the first release where Mistral and the Chinese open-weight flagships share one benchmark table.
| Benchmark | Mistral Large 4 | GLM-5.3 | DeepSeek V4.1 Flash | Kimi K3 |
|---|---|---|---|---|
| FrontierCode 2.0 | 19.7% | 17.8% | 19.3% | 17.3% |
| Terminal-Bench 4.0 | 41% | 40.0% | — | — |
| Artificial Analysis Cyber Index | 50.9% | 36% | 57% | 41% |
Numbers come from Mistral’s technical report, published November 3 on alphaXiv, and the announcement page. On Terminal-Bench, alphaXiv’s AI overview cautions the GLM-5.3 comparison was not a controlled run — the gap is thin enough to ignore. On FrontierCode, ML4’s lead over DeepSeek V4.1 Flash is 0.4 points, the thinnest margin in the table; on the Cyber Index the story flips outright: DeepSeek V4.1 Flash leads at 57% to ML4’s 50.9%, with MiMo-V2.6-Pro’s 56% also above it, though Mistral’s chart frames them as reference points rather than leaders being passed. The “leading open model” claim holds on the two coding boards, not on security. When we covered the October preview, we wrote that third-party reproduction after the weights landed would deliver the real verdict; the weights are here, the reproduction has barely started.
Community reaction and the license fight
The Hacker News thread hit 1,727 points with 440-plus comments, and the disagreement clustered in two places. First, the word “open”: several commenters argue a modified MIT that gates model-as-a-service is not OSI-compliant open source — “weights available” would be more accurate — a sharper contrast after Ant Group shipped Ling-3.1-flash under a standard MIT license. Second, value for money: one practitioner’s evals concluded the price cut “warrants a major reassessment,” while another asked about cache pricing — “my coding sessions have a >97% input cache rate. DeepSeek got us spoiled.” ML4’s $0.07 per million cached input tokens is an answer, though some still find the thinking mode slow and expensive for the output. The most representative dismissal came from vincentshipsit: “I pick a model on performance for the job, and from what I’ve seen, this isn’t it.” satvikpendem would reach for it only when EU provenance matters. Artificial Analysis’ independent page lists an Intelligence Index v5.0 of 26 — 10th of 23 in its cohort — a colder reading than the official “leading open model” framing.
Our take
For self-hosting teams, this is the first time in 2026 you can pick a trillion-parameter open-weight flagship from outside the US and China, and the 606GB deployment bar is almost a credibility signal — no small model dressed up as a release. For most developers, the economics stay awkward: ML4 leads DeepSeek V4.1 Flash by 0.4 points on the coding board while trailing it by 6.1 on security — margins a single reproduction could erase — DeepSeek’s cache pricing undercuts ML4’s, and the differentiation narrows to long context, multimodal input and European data residency. The real verdict waits on third-party reproduction of the agentic benchmarks — Terminal-Bench and the Cyber Index in particular, where Mistral’s self-built environment pipeline is both a strength and a potential scoring bias. Worth watching next: when community quantizers squeeze a usable build into a 4-GPU RTX server’s memory window, because that is the moment open weights stop being a declaration and become a fact.