Mistral launched a public preview of Mistral Large 4 on October 6, a mixture-of-experts model it describes as “1 trillion-parameter natively multimodal” with 52 billion active parameters, and said the weights will follow by the end of October. Mistral Today the model exists as an API on Mistral Studio with a published price, a docs page and a long benchmark post. The two things a self-hosting decision turns on, the checkpoint and its license, do not exist yet, and the independent record is one index score. The launch post names no license; the docs page tags the model “Open” where Mistral’s other models carry “Apache 2.0” or “Modified MIT”. Mistral docs, models
Update, October 30, 2026: The weights are out. Mistral published an FP8 checkpoint and an NVFP4 checkpoint on October 29, public and ungated, under a modified MIT license whose single condition applies to model-as-a-service businesses. Mistral Hugging Face model card The released model is not the October 6 preview: Mistral says it had “three weeks of additional reinforcement learning”. The update section below takes the October 7 checklist item by item; the article above it describes what was known on October 7, with dated notes where a statement no longer holds.
This is the same position Reflection’s Beam was in a day earlier, with one difference: Mistral has shipped large open checkpoints before, under more than one license. S5 Labs has not tested the model, and every benchmark below is Mistral’s unless marked as Artificial Analysis.
What you can use today
The preview is served through Mistral Studio and the standard endpoints: chat completions, conversations, agents and batch. The docs list structured outputs, function calling, document QnA and prefix completion, a 1M-token context window, and a 1.6B-parameter vision encoder. The model ID on the docs page is mistral-large-4 with one alias; the docs’ own compare links and OpenRouter’s listing use mistral-large-4-0. Mistral docs, Mistral Large 4 OpenRouter
Pricing per million tokens, from the docs page:
| List | Sale | |
|---|---|---|
| Input | $1.36 | $0.68 |
| Cached input | $0.14 | $0.07 |
| Output | $4.18 | $2.09 |
The sale is a straight 50% off with no end date on the page. Mistral docs, Mistral Large 4 For comparison, Mistral Large 3 lists at $0.50 input, $0.05 cached and $1.50 output, so Large 4 at list is about 2.7 times its predecessor on input and 2.8 times on output; at the sale price the gap is about 1.4 times. Mistral docs, pricing
October 30 note: the list and sale columns are gone. The docs now show one price, $0.70 input, $0.07 cached and $2.10 output, and the model is still marked Public Preview. See Pricing below.
Two things in the docs qualify the word “preview”. Mistral’s lifecycle policy says public preview models are “priced at the same rate as General Availability models”, can receive updates before GA, have “no guaranteed path to General Availability” and may be retired before reaching it, with a one-month deprecation notice instead of the six months a GA model gets. Mistral docs, model lifecycle And the context window is not consistent across Mistral’s own surfaces: the docs say 1M, while the OpenRouter endpoint served by Mistral lists 524,288 tokens with output capped at 262,144, and Artificial Analysis records 524k. OpenRouter Artificial Analysis Test the window you plan to use rather than the one on the model page.
1T or 1.05T
The launch post says 1 trillion total parameters. The docs page, versioned v26.10, says “52B active parameters and 1.05T total parameters, and a 1.6B vision encoder”. Mistral Mistral docs, Mistral Large 4 The post is probably rounding. Fifty billion parameters is a rounding error at this scale and 100 GB of BF16 weights when you are sizing a node, so the estimates below use 1.05T and show where 1T would change them.
Mistral has not published the layer count, expert count, experts per token or attention design. The post says those details, along with “additional benchmarks, and our post-training methodology”, will come with the weights. Mistral Our mixture-of-experts explainer covers why total and active counts answer different questions.
October 30 note: params.json is public. 72 layers, 128 routed experts with 4 active plus 1 shared, and a 1,048,576-token position limit; the model card gives 49B active per token, or 52B counting embeddings and output layers. Details in Architecture below.
The license is a blank badge, and Mistral’s record cuts both ways
Nothing Mistral has published for Large 4 names a license. The launch post uses “open weights” and “open-weight” repeatedly and never says under what terms. The docs page’s badge reads “Open”, which on the same models list sits next to “Apache 2.0” for Large 3 and Small 4, “Modified MIT” for Medium 3.5 and “Open” for Z.ai’s GLM 5.3, a third-party model with a custom license of its own. Mistral docs, models Artificial Analysis, which tracks weights availability, lists Large 4 Preview as proprietary, with weights not publicly available. Artificial Analysis On Hugging Face, the mistralai organization’s model list shows no Large 4 repository. Hugging Face, mistralai
The precedent is mixed. Mistral Large 3, released December 2, 2025, is Apache 2.0: the announcement says “All models are released under the Apache 2.0 license”, and the Hugging Face repository for Mistral-Large-3-675B-Instruct-2512 carries that tag. Mistral, Mistral 3 Hugging Face Mistral Medium 3.5, docs version v26.04, is not. Its LICENSE file is a “Modified MIT License” whose second clause says you are not authorized to use the model, or any derivative, if your company’s “global consolidated monthly revenue” exceeded $20 million in the preceding month, unless Mistral grants a commercial license or you use its hosted service. Hugging Face, Mistral Medium 3.5 LICENSE So within one year Mistral has shipped its flagship as plain Apache 2.0 and its next-largest model with a revenue gate. Large 4 could go either way, and the post gives no signal.
Recent open-weight launches show how much the term can cover. Qwen3.8-Max’s weights arrived under a custom license with a $50 million commercial threshold, and the checkpoint was a text-only core rather than the hosted product. GLM 5.3 split into an MIT Flash model and a flagship whose license sends model-as-a-service operators above $10 billion through a security review. Reflection’s Beam has promised Apache 2.0 with no text to read. Until a LICENSE file exists in a Large 4 repository, “open-weight” is a description of intent.
October 30 note: it went a third way. The LICENSE file is the “Mistral Ecosystem Open-Weight License”, a modified MIT with no revenue figure at all; its one condition turns on whether you operate a model-as-a-service business. See The license below.
Mistral’s benchmarks, and the one independent number
Mistral’s headline coding figures are DeepSWE v1.1 61.7%, SWE-Atlas-QnA 59.4% and Terminal-Bench 4 at 28.3%, with a Coding Agent Index of 49.8% that it says places the model ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max. A blind human evaluation run with Surge AI put “ML4 Preview” second of five at 3.74 on a five-point scale, behind Claude Opus 5 (4.22) and ahead of GLM-5.3 (3.60), Kimi K3 (3.59) and GLM-5.2 (3.40). On agents, Mistral cites 59.9% on AutomationBench and 1,393 Elo on AA-Briefcase; on vision, 42% on Dense 200 against 41% for GPT-6 Astra. An internal human evaluation against GLM-5.3 preferred Large 4 “in CAD and STEM, while performing on par or close to GLM-5.3 in finance and coding.” Mistral The post’s own claim is “competitive with the strongest open-source models globally” and “significantly outperforming any open-weight model developed in the US or Europe”. Its human eval has the model behind Opus 5, and the comparison set for the open claim is Chinese models it describes as roughly level.
Artificial Analysis had evaluated the preview by October 7. Its page gives Large 4 Preview an Intelligence Index of 38 (v4.3.2), above the median of 26 for comparable models, at 116 output tokens per second. Two of its figures matter for cost. The model generated 200 million tokens across the index run, “very verbose in comparison to the median of 81M”, and the resulting cost per index task is $1.13 at list price. Artificial Analysis A model that is cheap per token but verbose per task is the case our Sonnet 5.5 piece made for measuring cost per finished task. Mistral’s cyber claims rest on the same evaluator: it says the model “ranks among the top five models globally” on the Artificial Analysis Cyber Index and scores 82% on the index’s reproduce-and-patch test. We could not read the model’s cyber rank from the index page, whose text lists only the top three (Grok 4.7, MiMo-V2.6-Pro and GPT-6 Luna). Artificial Analysis, Cyber Index Treat the top-five claim as Mistral’s until the page confirms it.
October 30 note: the preview numbers above belong to the October 6 checkpoint. Artificial Analysis now lists the released model as a separate entry on Intelligence Index v5.0, a different composition from v4.3.2, so the 38 and the new 26 are not comparable. See The independent record below.
Cyber access is the same model with fewer refusals
The post is direct about why it leads with cybersecurity. It says Claude Opus 5.5 and GPT-6 Astra “score near zero” on the reproduce-and-patch test “because they refuse to perform the task”, and that defenders “need systems that can match those capabilities without being constrained by the same refusals.” Between now and the weights release, Mistral says it is “red-teaming the model in real-world settings with cybersecurity leaders, vetted partners, and state authorities, who will access the same model with reduced moderation and expanded cyber capabilities.” Mistral That is one model with two moderation settings, not a separate cyber build, and the post says the open weights will let organizations run security work “under their own policies”.
The same post reports a cyber refusal rate “higher than all OSS models” on JailbreakBench, StrongREJECT and AgentHarm prompts, so the public preview is tuned to refuse more than the partner build. Which behavior the released checkpoint carries is not stated. Anthropic’s answer to the same tension, published the same day, is a verification program that turns down cyber classifiers by tier for vetted applicants on hosted models; our coverage of that program is the useful comparison, because Mistral’s version of tiered access ends when the weights are public.
October 30 note: the released checkpoint is the guarded one, and the partner build is now described as a “specialized variant”, so the “one model, two settings” reading no longer matches what Mistral says it is doing. See Cyber below.
What a 1T, 52B-active checkpoint needs to run
No checkpoint exists, so these are estimates. Assumptions: decimal gigabytes, weights only, 1.05T parameters including the vision encoder, no KV cache, activations or runtime buffers. Precision figures use Mistral’s own Large 3 checkpoints as the ratio: the BF16 repository is 1,352 GB for 675B parameters (2.0 bytes per parameter), the default FP8 release 681.5 GB (1.01 bytes) and the NVFP4 build 403.1 GB, about 4.8 bits per parameter once scales are counted. Hugging Face, BF16 Hugging Face, FP8 Hugging Face, NVFP4
| Precision (estimate) | Weights at 1.05T | Weights at 1T | Published node or box that holds it |
|---|---|---|---|
| BF16, 2 bytes/param | ~2,100 GB | ~2,000 GB | Nothing in one node: DGX B300 lists 2.1 TB total GPU memory, which is the weights with no cache NVIDIA |
| FP8, ~1 byte/param | ~1,060 GB | ~1,010 GB | 8× H200 (1,128 GB) with roughly 70 GB spare NVIDIA |
| NVFP4, ~4.8 bits/param | ~630 GB | ~600 GB | 8× H200 comfortably; 8× H100 (640 GB) holds the weights with almost nothing left |
| Ideal 4-bit, 0.5 byte/param | ~525 GB | ~500 GB | A 512 GB Mac Studio M5 Ultra does not hold 1.05T at 4 bits; it would at 1T with no cache |
October 30 note: the estimates are superseded by published sizes. The FP8 repository is about 1.06 TB and the NVFP4 repository about 610 GB, so the FP8 and NVFP4 rows landed within a few percent; a BF16 checkpoint was not released, and the 4-bit GGUF is 635 GB, above the 512 GB a Mac holds. The published serving configurations are in Published configurations below.
The last row matters most to anyone pricing a desktop. Mistral said Large 3’s NVFP4 build runs “on a single 8×A100 or 8×H100 node”. Mistral, Mistral 3 Scaled by parameter count, a Large 4 NVFP4 build does not leave room for a cache on 8× H100 and needs the H200 generation or better. A 512 GB unified-memory desktop, which Apple has said ships in late October, is below the ideal 4-bit figure at 1.05T; running the model there means a quantization below 4 bits, with the quality loss that implies, and it would be a single-user box in any case. Four 128 GB DGX Spark units reach 512 GB on paper with the same shortfall plus an interconnect penalty. The Open Models guide separates file size from resident memory and the machine guide covers headroom by box.
Per generated token, the model reads its active parameters, about 104 GB at BF16 or roughly 31 GB at NVFP4, which is a bandwidth figure, not a capacity one, and is why a 52B-active model can be fast once it fits.
The KV cache is the unknown. Mistral has not published the attention design, and the cache size depends on it entirely. Large 3 uses multi-head latent attention with a 512-dimension latent plus a 64-dimension rotary component across 61 layers, which works out to about 70 KB per token at BF16, or around 18 GB for a 256K context and about 74 GB at 1M. Hugging Face, Large 3 params If Large 4 keeps that family of attention, full-context cache is tens of gigabytes per sequence; if it uses conventional grouped-query attention at this width, it could be several times more. Our KV cache explainer covers the arithmetic. Until params.json is public, size the node for the weights at your precision plus at least the Large 3 cache figure at your context, then add concurrency.
What to check when the weights land
- A
mistralai/Mistral-Large-4-*repository on Hugging Face, whether it is gated, and the LICENSE file’s text rather than the badge. Apache 2.0 puts it with Large 3; a modified MIT with a revenue clause puts it with Medium 3.5. Check that the metadata tag and the file agree. - Which checkpoint ships: instruct, base, or both; whether the released model carries the public preview’s moderation tuning or the partner build’s; whether the vision encoder is included. Qwen3.8-Max’s download was not the hosted product, and the launch post calls Large 4 the base for “a new generation of specialized and optimized Mistral models”, so read the card for what was held back.
params.json: total parameters against 1.05T, experts and experts per token, layers, attention type andmax_position_embeddings(1M or 524,288).- Whether BF16, FP8 and NVFP4 builds all appear at launch, as they did for Large 3, and whether the sizes land near the estimates above.
- The architecture and post-training notes the post promises, with eval settings for the headline numbers: effort, harness, pass@k, run counts, and the five-model set behind the Surge AI human eval.
- Independent entries: Artificial Analysis moving the model from proprietary to open weights and publishing its cyber rank; third-party hosts on OpenRouter beyond Mistral itself; a vLLM or SGLang recipe with a real memory footprint at a stated context.
- Whether the half-price preview rate survives GA, and the one-month notice window if the preview is retired or replaced.
Until the repository exists, Large 4 is a priced API with a strong vendor benchmark page and one independent index score. It is worth a preview-API evaluation now if your workload is coding, agents or document vision, with cost measured per finished task given the verbosity Artificial Analysis recorded. It is not yet something to plan hardware or a license review around, and a 1.05T checkpoint will not fit the hardware most teams have on hand at any precision Mistral has shipped before.
Update, October 30: what the weights answered
Mistral published the weights on October 29 under the heading “Mistral Large 4 (Le Chonk) is now open weight”. Mistral Two repositories are public and ungated on Hugging Face: mistralai/Mistral-Large-4-1T-A52B, the FP8 checkpoint, and mistralai/Mistral-Large-4-1T-A52B-NVFP4. Both were last modified on October 29; the FP8 repository was created on September 30, a week before the preview, and the NVFP4 one on October 15. Hugging Face, model card Hugging Face, NVFP4 card Hugging Face’s launch post carries the same date. Hugging Face blog
The post is explicit that this is not the October 6 model. Mistral says it is “releasing the weights of this improved Mistral Large 4” after “three weeks of additional reinforcement learning”, and puts its own Terminal-Bench 4.0 score at 41% against 28% for the preview, with 30% fewer output tokens. Mistral The docs page still reads Public Preview, October 6, 2026, v26.10. Mistral docs, Mistral Large 4 So a preview evaluation run three weeks ago measured a different checkpoint from the one on Hugging Face, and the preview’s lifecycle terms, one month’s notice and no guaranteed GA, still govern the API. The rest of this section takes the October 7 checklist in order.
The license: modified MIT, one condition, no revenue number
The LICENSE file in both repositories is the “Mistral Ecosystem Open-Weight License”, which calls itself “a modified version of the MIT license”. It grants the usual MIT rights to use, copy, modify, distribute, sublicense and sell copies, plus the right to “run, deploy, further train or otherwise modify the Model and create derivative works”. The second condition is the whole of the difference from MIT: “If the Licensee or any of its affiliates operates a Model as a Service business, the Licensee must enter into a separate commercial license with Mistral before using the Model or its derivative works for any commercial purpose.” LICENSE
The definition is the part to read twice. “Model as a Service” means “giving a third party access to the Model or its derivative works for inference or post-training by any means (for instance via API or hosted endpoints), in a manner that allows such third party to exercise meaningful control over the inputs, parameters, or training data.” Two carve-outs follow: “end-user products with model capabilities solely embedded within specific features or harnesses”, and “mere relaying of requests to models hosted by others”. A third clause exempts internal use, “any use that does not make the Model or its derivative works available to third parties”, and any use through Mistral’s official products. The contact route for a commercial license is a web form. LICENSE The model card and the NVFP4 card summarize it the same way, and the docs badge that read “Open” on October 7 now reads “Mistral Ecosystem Open Weight License”. Hugging Face, model card Mistral docs, models The Hugging Face metadata tag is license: other, so the file, not the tag, is the record.
So the October 7 framing, Apache 2.0 like Large 3 or a revenue gate like Medium 3.5, was the wrong pair; there is no revenue threshold. The test is a business model: a company of any size that runs the model internally, or ships it inside a product feature, is clear; a company of any size that sells access to the model itself, or fine-tunes it for third parties, needs Mistral’s paper first. Meta’s Muse Spark Open license has the same shape with a $20 million revenue floor under it; Mistral removed the floor. That makes Large 4 stricter than Muse Spark Open for a small hosting provider and looser than Medium 3.5 for a large enterprise with an internal deployment. The gray zone is a developer platform whose customers bring their own prompts and sampling parameters, which sounds like “meaningful control over the inputs, parameters” even if the model is “embedded” in a product; the license does not resolve that, and the form is where you ask. Artificial Analysis now files the model as open weights with the license named on the page. Artificial Analysis
What shipped, and what did not
One checkpoint, not an instruct and base pair. The card describes “instruct and reasoning functionalities” in a single set of weights with a per-request reasoning effort of low, medium or high, defaulting to high. Hugging Face, model card Reasoning cannot be turned off: the llama.cpp card says so, and Hugging Face’s post adds that passing none still runs high. ggml-org GGUF card Hugging Face blog The vision encoder is included: a 1.6B Pixtral encoder with 30 layers, shipped as a separate mmproj file in the GGUF repositories. Hugging Face blog ggml-org GGUF card
The main repository is FP8, not BF16. Its config.json sets quant_method: fp8 with 128×128 weight blocks, and vLLM’s recipe calls it “FP8 block-scaled, ~1.06 TB”. vLLM recipe Mistral released no BF16 checkpoint; the 2.10 TB BF16 file in the llama.cpp repository is described as a “lossless upcast of the FP8 release”. ggml-org GGUF card That differs from Large 3, which shipped BF16, FP8 and NVFP4 together.
The checkpoint also ships something the preview did not advertise: an embedded DSpark speculative-decoding draft head, three dense MLA draft layers in BF16 that propose blocks of three tokens. Mistral says it measured up to 1.9 times the throughput of serving the target model alone, depending on hardware and workload. Hugging Face, model card Both the vLLM and SGLang examples turn it on with a flag.
Architecture: what params.json says
The October 7 article could only guess from Large 3. The params.json in the NVFP4 repository, and the model card, answer most of it. params.json Hugging Face, model card
| Item | October 7 question | Answer |
|---|---|---|
| Total parameters | 1T or 1.05T | Card says “~1T”; docs still say 1.05T; the FP8 repository stores 1.06 TB, which at about a byte per parameter supports the docs figure |
| Active parameters | 52B | 49B per token, 52B “including embeddings and output layers” (card). The repository name keeps A52B; the October 7 figure stands |
| Experts | not published | 128 routed, 4 active per token, plus 1 shared expert; the first layer is dense |
| Layers | not published | 72, hidden size 6,144, 64 attention heads |
| Attention | MLA like Large 3, or GQA | MLA with a 512-dimension latent and 64-dimension rotary component, as in Large 3, plus a DeepSeek-style sparse indexer that selects up to 2,048 keys per query; KV state is shared across some layers |
max_position_embeddings | 1M or 524,288 | 1,048,576 |
| Context in practice | vLLM validated to 256K; Hugging Face’s endpoint recipe serves 131K; OpenRouter’s listing now shows 1,048,576 with 262,144 output tokens vLLM recipe Hugging Face blog OpenRouter |
The 524,288 figure on OpenRouter was an endpoint setting; the checkpoint allows 1,048,576 positions. The 49B-versus-52B split is accounting, not a correction: both numbers describe the same checkpoint, and the docs, the repository name and Artificial Analysis use 52B.
Published configurations: what Mistral, vLLM and Hugging Face actually ran
The October 7 table was arithmetic. These are the configurations someone has run. Sizes are the repositories’ storage figures and the recipe’s stated weights per GPU. S5 Labs has run none of them.
| Checkpoint | Size | Published configuration | Source |
|---|---|---|---|
FP8, mistralai/Mistral-Large-4-1T-A52B | ~1.06 TB | Model card examples: vLLM at tensor parallel 8, SGLang at --tp-size 8, described as “example tensor-parallel size for a node of 8 GPUs; adjust to your topology”. vLLM’s recipe validated FP8 on 8× B300, 8× GB300 and 2×8 H200 (TP16 or PP2×TP8) and gives “~128 GiB of weights per GPU on 8 GPUs”. Hugging Face’s Inference Endpoints recipe runs it on 8× H200 with tensor and expert parallelism, 131K context, 16 concurrent requests | Model card, vLLM recipe, Hugging Face blog |
NVFP4, mistralai/Mistral-Large-4-1T-A52B-NVFP4 | ~610 GB | Blackwell only: the card says it “targets Blackwell GPUs (B200 / B300 / GB200 / GB300), which have dedicated hardware for the format”. SGLang example at --tp-size 4; vLLM example at tensor parallel 8; vLLM’s recipe default is 4× GB300 NVL4 at TP4, with “~76 GiB” of weights per GPU on 8 | NVFP4 card, vLLM recipe |
AMD MXFP4, amd/Mistral-Large-4-1T-A52B-Quark-MXFP4 | ~578 GB | MI350 series; vLLM says the MXFP4 experts fit four GPUs, so an 8-GPU node serves two TP4 replicas | vLLM recipe |
| GGUF Q8_0 / Q4_K_M | 1.12 TB / 635 GB | Hugging Face: these “require distributed inference” | ggml-org GGUF card, Hugging Face blog |
| GGUF Q2_K / IQ2_XXS | 381 GB / 298 GB | “Fits in 512 GB of unified memory”; Hugging Face says the 2-bit quants run on a single Apple Silicon Ultra with 512 GB, and reports about 90 tokens per second prefill and 15 tokens per second generation for Q2_K on an M3 Ultra | ggml-org GGUF card, Hugging Face blog |
First, eight GPUs is the reference setup, not a floor. Every example from Mistral uses tensor parallel 8 on the FP8 checkpoint, but the NVFP4 card’s SGLang example and vLLM’s default both use four Blackwell GPUs, and AMD’s MXFP4 build fits four MI350s. NVFP4 card vLLM recipe For FP8 specifically, about 1.06 TB of weights across eight 141 GB H200s leaves roughly 70 GB for cache and runtime, which is our arithmetic and the same number the October 7 table reached; vLLM validated FP8 on sixteen H200s rather than eight, while Hugging Face squeezed eight by capping context at 131K and concurrency at 16. vLLM recipe Hugging Face blog Treat 8× H200 as the FP8 minimum that works with a short context, not a comfortable one.
Second, the Hopper-versus-Blackwell split is now explicit. NVFP4 needs Blackwell’s hardware support, so the 610 GB build is not an H200 option in the vendor’s own recipe; vLLM’s validation list for NVFP4 is GB300, B300 and B200 only. NVFP4 card vLLM recipe An 8× H200 node runs FP8 or nothing from Mistral; the smaller GGUFs are the other route, and Hugging Face reports an NVFP4 GGUF on four RTX 6000 Pro cards at about 10 tokens per second generation with a 32K context. Hugging Face blog
Third, the 512 GB desktop question is settled in the direction the October 7 piece expected. The 4-bit GGUF is 635 GB and does not fit; the 2-bit ones do, and IQ2_XXS reaches 298 GB by quantizing only the routed experts to 2 bits while keeping attention, the shared expert, the dense first layer, embeddings and the output head at Q8_0. ggml-org GGUF card Hugging Face’s 15 tokens per second on an M3 Ultra is a single-user figure; at long contexts its four-GPU IQ2_XXS run fell to 3 tokens per second generation. Hugging Face blog Nobody has published a quality comparison of the 2-bit files against FP8, and the NVFP4 card’s “no accuracy regressions” claim covers NVFP4 only. NVFP4 card Our machine guide covers what a 512 GB Mac is and is not for.
KV cache: now an arithmetic problem, not an unknown
With params.json public, the cache can be sized. Each layer’s MLA cache holds a 512-value latent plus a 64-value rotary component per token, the same layout as Large 3, and the checkpoint sets an FP8 KV cache; vLLM uses its FP8 latent cache for this model by default. params.json vLLM recipe At one byte per value that is 576 bytes per token per layer. params.json also lists cross-layer KV sharing in which nine of the 72 layers reuse one earlier layer’s cache, leaving 63 layers with caches of their own. The product is about 36 KB per token, or roughly 9 GB for a 256K sequence and 36 GB at the full 1M, before the sparse indexer’s own cache and the draft head. That is S5 Labs arithmetic from the published file, not a figure Mistral or vLLM has stated, and it is about half the Large 3 BF16 figure per token because the cache is FP8. The number that is published: vLLM’s recipe reports 3.60M cached tokens on 8× GB300 at tensor parallel 8 with NVFP4 weights, rising to 28.76M with decode context parallel 8, “~27 concurrent 1M-token requests by KV size”, with the recipe noting that serving at 1M context is not validated. vLLM recipe The eightfold gap is what you would expect if plain tensor parallel replicates the MLA cache on every GPU, which would explain why the long-context presets use context parallel. Our KV cache explainer covers why MLA caches do not split by head.
Pricing: the sale rate became the rate
The news post says the model “is priced at $0.70 input / M, $0.07 cached input / M, and $2.10 output / M tokens” from October 29, and the docs and pricing pages show those numbers with no list or sale column. Mistral Mistral docs, pricing
| Per 1M tokens | October 7 list | October 7 sale | October 29 |
|---|---|---|---|
| Input | $1.36 | $0.68 | $0.70 |
| Cached input | $0.14 | $0.07 | $0.07 |
| Output | $4.18 | $2.09 | $2.10 |
Against the October 7 list price this is a cut of about half. Against the sale price it is a rounding up of one or two cents, and the sale price is what anyone evaluating the preview was paying. The docs no longer show the $1.36 and $4.18 figures, so the sale did not end; the list price did. At $0.70 and $2.10 the model sits at 1.4 times Large 3 on both input and output. Mistral docs, pricing The October 7 question, whether the half-price rate survives GA, has a partial answer: GA has not happened, the docs still say Public Preview, and the lifecycle policy prices previews at GA rates, so the $0.70 rate is the standing one until Mistral changes it, with the preview’s one-month notice if it does. Mistral docs, model lifecycle The docs’ deprecation table now names Large 4 as the alternative for Large 3 and Medium 3.5, both with deprecation dated October 31, 2026 and retirement April 30, 2027. Mistral docs, models
The independent record after release
Artificial Analysis now lists the released model as “Mistral Large 4 (High)”, an open-weights model with the license named. Its page shows an Intelligence Index of 26 on index v5.0, ninth of 22 on its summary card, against a median of 23 for comparable models (the page does not name the peer group); 110M output tokens across the index run against a median of 99M; $2.06 per index task at the new prices; and 161 tokens per second. Artificial Analysis None of those compare with the October 7 figures. The 38 was on index v4.3.2 and the 26 is on v5.0, which names a different set of nine evaluations, so the drop is a change of ruler, not evidence the model got worse. The cost per task nearly doubled, from $1.13 to $2.06, for the same reason: a different task mix, priced at a different rate.
Rankings depend on who is counting. The model card says 26 is “the highest among western open-weight models”. Hugging Face, model card Hugging Face’s post says the model is “the fifth best open weights model on Artificial Analysis Intelligence index” as of October 29. Hugging Face blog Artificial Analysis’s own page says ninth of 22 on its summary card. Those are three different cuts of one table, and only the last is the evaluator’s. The Hugging Face post’s coding section also repeats the October 6 preview numbers, 61.7% on DeepSWE v1.1, 28.3% on Terminal-Bench 4 and a 49.8% Coding Agent Index, while Mistral’s own release post reports 48% on the Coding Agent Index and 41% on Terminal-Bench 4.0 for the released checkpoint. Hugging Face blog Mistral
The new vendor benchmarks, all Mistral’s: 19.7% on Cognition’s FrontierCode 2.0, which it says leads the open-weight set; 48% on the Coding Agent Index, “effectively tied” with GLM-5.3 at 47% fewer tokens per task; Surge Code blind pairwise preference over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, with Claude Opus 5.5 the only evaluated model ranked higher; 73.3% on FinWorkBench; 84.5% on BrowseComp; 97.1% attack resistance on the B3 agent-security benchmark against 93.3% for the preview. Mistral The Surge set has changed shape: the preview’s five-model, five-point scale eval is now a four-opponent win-rate eval, so the two cannot be stacked. The post promises “a full technical report” and does not give effort settings, run counts or pass@k for any figure, so the October 7 item on eval settings stays open.
Cyber: the public checkpoint is the guarded one
The October 7 reading was one model with two moderation settings. The release post describes something else. The public checkpoint refuses: Mistral reports an average refusal rate of 96.6% on the cyber prompts of JailbreakBench, StrongREJECT and AgentHarm, “the highest we measure among open models”, and the card says the model does security work “while refusing malicious requests”. Mistral Hugging Face, model card Mistral also says that in live preview traffic it “observed the model refusing to comply with black-box attacks against live targets”, that it sees “a potential for uplift to conduct real attacks”, and that it is “modifying our guardrails accordingly” with updates to follow. The partner build is now “a specialized variant of Mistral Large 4 optimized for advanced cybersecurity workflows, including penetration testing, vulnerability assessment, and red-teaming”, provided to design partners such as Orange Cyberdefense under vetting and contractual safeguards. Mistral That variant is not in the public repository. Whether it is the same weights with different post-training or a different checkpoint is not stated.
On the index itself, Mistral now gives a number: 51% on the Artificial Analysis Cyber Index. Mistral Hugging Face repeats Mistral’s “among the top five models globally”; the index page’s text does not state a rank. Artificial Analysis, Cyber Index Hugging Face blog Hugging Face’s post repeats the preview’s 82% reproduce-and-patch figure; Mistral’s release post does not restate it. Since the public weights refuse at the rate Mistral reports, the October 7 point stands in a modified form: open weights put the usage policy in the operator’s hands, but the checkpoint Mistral shipped was tuned to refuse more, not less, and the unguarded capability lives in a variant you cannot download.
What is still open
- The technical report, with eval settings for every headline figure.
- Whether the hosted providers now carrying the model, Microsoft Foundry, NVIDIA’s NIM endpoints, Hugging Face Inference Providers and the third-party endpoints on OpenRouter, hold the commercial license the MaaS clause requires. The posts list the hosts; neither says. Mistral Hugging Face blog For your own deployment the question is narrower: does anyone outside your organization get to control inputs, parameters or training data.
- A published memory footprint at 1M context from anyone other than vLLM’s KV-capacity figure, and a quality comparison for the 2-bit GGUFs.
- GA, and with it the six-month deprecation notice that the preview label withholds.
What to do with this
Teams with an 8× H200 node can run FP8 today at a short context, and should budget for 16 GPUs or a Blackwell tray before promising long contexts. Teams buying Blackwell get the 610 GB NVFP4 build on four GPUs with vLLM’s default recipe, which is the configuration the vendor and the serving framework agree on. A 512 GB Mac is a 2-bit, single-user box at 15 tokens per second by Hugging Face’s measurement, useful for private evaluation and not for serving. For everyone else the API at $0.70 and $2.10 is the rational default, measured per finished task, where Artificial Analysis’s $2.06 on its v5.0 index is the first independent figure. The license review is short if you are honest about your business: internal use and embedded features are clear, selling model access or fine-tuning for third parties is not, and the absence of a revenue threshold means being small does not help. The October 7 judgment, that this was not yet something to plan hardware or a license review around, has expired; both are now plannable, on published terms.
Key Details
As of October 7, before the weights shipped.
| Spec | Detail |
|---|---|
| Lab | Mistral AI (Paris) |
| Announced | October 6, 2026; public preview API on Mistral Studio |
| Weights | Promised “end of this month”; no Hugging Face repository as of October 7 |
| License | Not named; docs badge “Open”. Large 3 was Apache 2.0; Medium 3.5 was modified MIT with a $20M monthly-revenue gate |
| Total parameters | 1T (launch post) or 1.05T (docs, v26.10), self-reported |
| Active parameters | 52B per token (about 5%) |
| Vision encoder | 1.6B (docs only) |
| Context | 1M (docs); 524,288 with 262,144 max output on the OpenRouter endpoint served by Mistral |
| Model ID | mistral-large-4 (docs); mistral-large-4-0 on OpenRouter |
| Price (per 1M tokens) | List $1.36 in, $0.14 cached, $4.18 out; sale $0.68 / $0.07 / $2.09, no end date |
| Preview terms | Same price as GA, updates possible, one-month deprecation notice, no guaranteed GA |
| Training | 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s European datacenters; 160+ languages (self-reported) |
| Vendor benchmarks | DeepSWE v1.1 61.7%; Terminal-Bench 4 28.3%; Cybench 93% of 40; Dense 200 42%; Surge AI human eval 3.74, second of five |
| Independent | Artificial Analysis Intelligence Index 38 (v4.3.2), 116 tokens/s, $1.13 per index task, 200M output tokens in the run |
| Self-host (estimates) | ~2,100 GB BF16; ~1,060 GB FP8; ~630 GB NVFP4; KV cache unknown until the architecture is public |
Key Details as of October 30
| Spec | Detail |
|---|---|
| Weights released | October 29, 2026; mistralai/Mistral-Large-4-1T-A52B (FP8, ~1.06 TB) and mistralai/Mistral-Large-4-1T-A52B-NVFP4 (~610 GB); public, ungated |
| License | Mistral Ecosystem Open-Weight License: modified MIT; model-as-a-service businesses need a separate commercial license; internal use and embedded product features exempt; no revenue threshold |
| Checkpoint | Single instruct-and-reasoning model, reasoning effort low/medium/high, cannot be disabled; vision encoder included; embedded DSpark draft head; no base model, no BF16 release |
| Architecture | ~1T total (docs: 1.05T); 49B active, 52B with embeddings and output layers; 72 layers; 128 routed experts, 4 active, 1 shared; MLA with sparse indexer; max_position_embeddings 1,048,576 |
| Published configs | FP8: TP8 examples, 8× H200 (Hugging Face Endpoints, 131K context) or 2×8 H200 and 8× B300/GB300 (vLLM). NVFP4: Blackwell only, 4× GB300 (vLLM default) or TP4/TP8 examples. GGUF 2-bit (298–381 GB) on a 512 GB Apple Silicon Ultra |
| KV cache | ~36 KB per token at FP8 by S5 Labs arithmetic from params.json; vLLM reports 3.60M cached tokens at TP8 on 8× GB300, 28.76M with decode context parallel 8 |
| Price (per 1M tokens) | $0.70 in, $0.07 cached, $2.10 out; still Public Preview, v26.10 |
| Vendor benchmarks (release post) | FrontierCode 2.0 19.7%; Coding Agent Index 48%; Terminal-Bench 4.0 41%; AA Cyber Index 51%; FinWorkBench 73.3%; BrowseComp 84.5%; B3 97.1%; cyber refusal 96.6% |
| Independent | Artificial Analysis Intelligence Index 26 (v5.0), 9th of 22 on its summary card, 110M output tokens, $2.06 per task, 161 tokens/s |
| Still pending | Technical report and eval settings; GA; quality data on 2-bit GGUFs; the cyber “specialized variant” is partner-only |
Sources
- Introducing Mistral Large 4 — Mistral AI
- Mistral Large 4 (Le Chonk) is now open weight — Mistral AI, October 29, 2026
- Mistral Large 4 model page — Mistral docs
- Models overview — Mistral docs
- Pricing — Mistral docs
- Model lifecycle policy — Mistral docs
- Mistral-Large-4-1T-A52B model card, LICENSE, NVFP4 card and params.json — Hugging Face
- Welcome Le Chonk: Mistral Large 4 on Hugging Face — Hugging Face, October 29, 2026
- ggml-org/Mistral-Large-4-1T-A52B-GGUF — Hugging Face
- Mistral-Large-4-1T-A52B recipe — vLLM Recipes
- Introducing Mistral 3 — Mistral AI
- Mistral-Large-3-675B-Instruct-2512, BF16 and NVFP4 — Hugging Face
- Mistral-Medium-3.5-128B LICENSE — Hugging Face
- mistralai organization — Hugging Face
- Mistral Large 4 — Artificial Analysis
- Artificial Analysis Cyber Index — Artificial Analysis
- mistralai/mistral-large-4-0 — OpenRouter
- H200 Tensor Core GPU and DGX B300 — NVIDIA