On October 6, 2026, Mistral announced Mistral Large 4 (nicknamed Le Chonk) in public preview, with the API model ID mistral-large-4 open to all users that day. The spec sheet arrived fast and full: 1.05T total parameters, 52B active per token — or 49B excluding the embedding and output layers, more on the two figures below — plus a 1.6B vision encoder, 160+ languages of training data, and 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European data center, trained from scratch.
But as of press time (Oct 23), the promised open weights have not landed. The Hugging Face repo is a countdown placeholder with zero downloadable files and a license field reading "coming soon." The official line is one sentence: weights arrive by the end of the month. Repo: Mistral-Large-4.0-1T05-A52B, plain address: huggingface.co/mistralai/Mistral-Large-4.0-1T05-A52B — currently a placeholder. What "open" means for a model with no downloadable weights is the question this article answers.
A note on sources: Mistral's official pages could not be fetched directly in our environment, so official figures were cross-checked against at least three independent secondary sources; vendor and independent scores are kept apart.
One. Release Timeline: Specs Early, Weights Later
October 6: announcement and API opening the same day. October 21: at GTC Berlin, Mistral and NVIDIA announced an industrial engineering AI alliance covering CAE, computational physics, and EDA, built on CUDA-X, Warp, and Omniverse, reconfirming Large 4's architecture (roughly 1T total, 49B active, 3,800 Grace Blackwell GPUs) and the "open weights by end of October" schedule, plus co-development of open-weight frontier models inside NVIDIA's Nemotron Coalition and a compute roadmap pointing at 1 GW by 2030. Two confirmations in half a month: the pledge is not a marketing slip — it just is not a delivery yet.
The background matters too. Behind this release sits a EUR 3B Series D, the largest equity round for a European tech company, led by Samsung at a EUR 21B valuation after a Series C led by ASML. Science VP Pierre Stock told TechCrunch the GPU count was 2-3x smaller than Chinese competitors' (secondhand attribution). And the RL scale: a single run on roughly 3,000 GPUs produced about 33 billion rollout tokens per day, filtered to about 16 billion trainable tokens, reward still climbing — the Oct 6 scores are a snapshot of training in progress.
The go-to-market story is European end-to-end deployment: independently of other digital service providers and under European law, with private-cloud and on-premises options promised. Before launch, Mistral ran red-teaming with cybersecurity vendors, vetted partners, and state authorities — the latter two get a less-filtered version with expanded cyber capabilities.
Two. The Parameters-and-Pricing Ledger
First accounting issue — active parameters. The docs model card says 52B, and the HF repo name Mistral-Large-4.0-1T05-A52B encodes A52B. Artificial Analysis records 49B. Both are correct: 52B includes the embedding and output layers; 49B is the per-token forward pass without them, and the 3B gap is exactly those layers. Same with totals: the announcement rounds to "1-trillion-parameter," the model card says 1.05T. Dividing 49B by 1,050B gives an active ratio of about 4.67% (third-party calculation). The architecture is a granular MoE with hybrid instruct-and-reasoning behavior, but expert count, routing top-k, and layer count are unpublished — they ship with the weights.
Second issue — context. The official card lists 1M tokens, the API/hosted figure. OpenRouter's routing and Artificial Analysis's test configuration both use 524,288, i.e., 512K. The numbers coexist; cite the source when quoting either. The checkpoint's native context waits for config.json.
Pricing is where it gets interesting. List price: $1.36/M input, $0.14/M cached input, $4.18/M output. The docs page shows a struck-through half-price column on the same screen: $0.68, $0.07, $2.09, no stated end date. The discount's nature has three competing descriptions — launch rate, 50% batch discount, sale price — though all sources confirm both sets of numbers exist on the docs page. Best practice: quote list price and note the preview-period half price — when the discount ends, the bill doubles. The comparison is unforgiving: on OpenRouter this price is roughly 10x DeepSeek's.
Three. Auditing the Word "Open"
The heart of this article: taking "open" apart and inspecting the goods, verified across multiple sources as of Oct 22. The HF repo is a countdown page with no downloadable files, a license field reading "coming soon," and a README with no context length; Mistral's HF organization page showed no Large 4 checkpoint as of Oct 13. The docs side mirrors it: the license badge renders a single word, "Open," with no license name; in the model selection guide, Large 4's License cell is blank while neighboring rows are complete — Mistral Small 4 is Apache 2.0, Mistral Medium 3.5 is Modified MIT, Mistral Large 3 is Apache 2.0. Artificial Analysis's independent record: license name and URL both null, isOpenWeights false, categorized as proprietary — in AA's words, not a denial of Mistral's marketing, but a statement of what exists today.
Now the promise. The official wording is "We will release the weights by the end of the month," with architecture details, more benchmarks, and post-training methodology promised alongside. The HF placeholder lists an ETA of Oct 31; as of press time (Oct 23) eight days remain, and the weights could land any day — but today they have not.
The dividing line sits at the license. If the intent were Apache 2.0, the simplest move would be to name it; media coverage broadly reads the silence as license uncertainty, possibly custom commercial terms (secondhand interpretation). History supports the optimistic reading: Large 3 shipped as Apache 2.0 with FP8, NVFP4, and BF16 checkpoints downloadable. But a license is judged by the release in front of you, and until the field changes from "coming soon" to an actual name, "open weights" is expectation, not fact. Artificial Analysis's proprietary classification follows the same logic: only the API exists today — architecture confirmed, schedule confirmed, alliance signed, but "open" stays future tense until everything lands.
Four. Benchmarks: Two Very Different Pictures
Vendor-reported scores (launch blog; no independent full replication as of press time):
| Benchmark | Score |
|---|---|
| DeepSWE v1.1 | 61.7% |
| Terminal-Bench 4.0 | 28.3% |
| SWE-Atlas-QnA | 59.4% |
| Coding Agent Index (composite) | 49.8% |
| Surge AI blind eval (human, 1-5) | 3.74, second of five |
| AutomationBench | 59.9% |
| AA-Briefcase | 1,393 Elo |
| Cybench | 93% (37 of 40 solved) |
| Lakera B3 (prompt-injection resistance) | 93.3% |
| Dense 200 (visual grounding) | 42%, vs GPT-6 Astra 41.5% |
Details worth a closer look. The Surge AI blind eval of 3.74 lost only to Claude Opus 5 (4.22) while beating GLM-5.3 (3.60) and Kimi K3 (3.59) — human blind evals carry weight, but the sample and process are unpublished. The AutomationBench wording must be precise: the official text claims wins over Kimi K3, MiMo-V2.6-Pro, and DeepSeek V4 Pro — never over GLM-5.3, because Mistral's own chart shows GLM-5.3 at 62.2, above 59.9; calling it "best open model" would be wrong. The security self-reports show tension too: on the AA Cyber Index reproduce-and-patch task, Large 4 scored 82%, claimed as the field's best, while the same blog says its malicious-cyber prompt refusal rate is higher than any open model's. Both hold only because competitors refuse heavily — third-party data shows refusal rates of 98.5% for Opus 5.5 and 100% for GPT-6 Astra on the same task — so much of that 82% first place is policy gap, not capability gap. Dense 200's 42% versus 41.5% is the same story: chart reads as a win, numbers read as a tie. The official DeepSWE leaderboard (updated Sep 22) has not admitted Large 4, so 61.7% has one source only.
Independent third-party scores (around Oct 15) paint a different picture:
| Metric | Score |
|---|---|
| AA Intelligence Index | 38 (Opus 5.5 = 58, GPT-6 Astra = 53; similarly priced reasoning models' median = 26) |
| AA Omniscience (hallucination-penalized knowledge test) | -5.3 (Kimi K3 = 19.7, GLM-5.3 = 14.3) |
| AA factual-recall hallucination rate | 41.9% |
| AA MMMU-Pro | 76.4 (Kimi K3 = 80.5, Opus 5.5 = 87.7) |
| AA Cyber Index | 50, fifth worldwide |
| Vals Index | 48.05%, 33rd of 45 (Opus 5.5 = 66.97%) |
Artificial Analysis's positioning line — "the most intelligent model from outside the US and China" — can be read positively, or as the other half of the same sentence: 38 points is 20 behind the frontier. Simon Willison's first look also landed at 38, concluding the model trails the frontier by roughly six months, with high reasoning effort producing fewer output tokens on his pelican test; extrapolated from the Vals leaderboard, the gap is about 9-12 months. The -5.3 on Omniscience is negative — bare-recall fabrication is worse than admitting ignorance, and the 41.9% hallucination rate confirms it: a model for RAG and document-grounded work, not bare recall. On speed, AA recorded 116 tok/s; OpenRouter measured 70-80 tok/s on Oct 15.
Five. Community Reaction: Heat, Doubt, Defense
The main Hacker News thread hit 1,449 points "before dinner" on launch day, finishing at 1,826 points with 1,085 comments — the week's hottest model release; some outlets logged two threads at 613 and 319 points, unverified. Skepticism clustered in two places. On capability, a commenter ran private adversarial coding benchmarks and found the performance "quite poor — about the same level as Kimi K3." On price, the roughly 10x-DeepSeek figure on OpenRouter drew the highest-frequency rebuttal: "why would I use this, then?"
The defense line is EU data residency: for European enterprises, defense, and critical infrastructure, GDPR and sovereign deployment are real reasons — but not for GPU-holding hobbyists. The consensus on the open-weights landscape: GLM-5.3, MiMo, and Kimi K3 remain the Pareto frontier; ML4's overall capability is close to GLM-5.3 and comparable on cyber, but below K3 and MiMo on broad benchmarks; DeepSeek V4.1 Flash wins on price-performance.
Chart criticism used the phrase "chart-crime adjacent": bar ordering places Mistral next to the weakest competitor, some axes do not start at zero, some competitor scores have no public source. On security there was a same-day counterpoint: Anthropic expanded its Cyber Verification Program into a three-tier vetted-access scheme — open weights versus tiered authorization, two answers to the same double-edged question. The branding drew near-zero criticism — Le Chonk was praised for not taking itself too seriously. One pragmatist comment summed it up: compare on your own tasks, not the vendor's chart; the bill doubles when the discount ends; RL is still running, so Oct 6 scores are a moving target.
Six. A Cold Thought: The Local-Run Memory Threshold
Suppose the weights do land by end of October — who can run them? Third-party estimates (not official): about 2.1TB in BF16, 1.05TB in FP8, 525GB at 4-bit. The threshold sits at 400GB-plus of fast memory; today the only single machine that theoretically qualifies is the Mac Studio M5 Ultra with 512GB of unified memory. This is the same story as Beam 501B, which needs four B200s to carry its 1TB of weights — open weights at this tier were never a desk-side product.
The whole article in one paragraph: Mistral proved its engineering and fundraising strength with a release generous in specs and narrative, then reminded everyone with seventeen days of missing weights and an unnamed license that "open" is a future-tense adjective here. The end of October will settle it — until then, Large 4 is a proprietary model behind a pricey API.
FAQ
Q1: Can I download the Mistral Large 4 weights now?
A: Not as of press time (Oct 23). The HF repo is a countdown placeholder with zero files and a "Coming soon" weights tab; the official pledge is "by the end of the month," and the HF placeholder lists an ETA of Oct 31 — check the repo's live status.
Q2: Is it actually open source? What license?
A: Undetermined, for now. The docs badge renders only "Open," the HF field reads "coming soon," and Artificial Analysis classifies it as proprietary because the weights are not downloadable. Mistral Large 3 shipped as Apache 2.0, which supports the optimistic reading — but until a license is named, openness is expectation, not fact.
Q3: Is it 52B or 49B active parameters?
A: Both, under different accounting. 52B includes the embedding and output layers — the figure in the HF repo name A52B; 49B is the per-token forward pass without them, the figure Artificial Analysis records. The 3B difference is exactly those layers.
Q4: Is the context window 1M or 512K?
A: Both numbers coexist. The official card lists 1M tokens (API and hosted figure); OpenRouter routing and Artificial Analysis test configurations both use 524,288 (512K). The open-weights checkpoint's native context awaits config.json.
Q5: Can I run it locally once the weights ship?
A: Almost certainly not on consumer hardware. Third-party estimates put BF16 at about 2.1TB, FP8 at 1.05TB, and 4-bit at 525GB; the threshold is 400GB-plus of fast memory, and only the 512GB unified-memory Mac Studio M5 Ultra theoretically qualifies today. Otherwise there is the API at $1.36/$0.14/$4.18 per million tokens list price, half price during the preview.
Related reading
- Reflection AI Beam 501B: Can Four B200s Carry 1TB of Weights?: the same announce-first-ship-later script, kept in eleven days
- DeepSeek Weights Are Free, Yet Agent Task Bills Fall 60 Percent: the reference point in HN's price complaints
- 42B Active, 816B to Load: Why It Still Needs Two Macs: the desktop-grade version of the open-weights route
- Says #1, Its Own Table Says #2: Picking a 24GB Model: more cases where vendor charts and measurements diverge