AI Models

Mistral Large 4 Open Weights: The 1T Deployment Reality

Mistral Large 4's October 29 weight release changes the open-model frontier. Audit its 1T/52B MoE design, FP8 and NVFP4 costs, sparse attention, benchmarks and licensing.

Approximately 8 min read

On October 29, 2026, Mistral published the weights of Mistral Large 4, nicknamed Le Chonk. This is not merely another API model announcement. An October 6 public preview already existed; the new event is the availability of improved, downloadable weights, an official FP8 checkpoint, an NVFP4 variant, a stated commercial-use license and deployment instructions. The distinction matters: a model can be impressive in a hosted demo yet impractical to operate on a single workstation.

The headline specification is roughly one trillion total parameters with 49 billion active in the core language model per token, or about 52 billion including embedding and output components. Its advertised context window is one million tokens, with text and image input and text output. The sparse expert count is about computation, not a promise that only 52 billion parameters need to reside in memory.

This article checks the Mistral announcement, official model artifacts and Hugging Face’s release analysis. All performance figures are externally reported; no local inference, download or independent benchmark was performed.

What shipped on October 29, rather than October 6

Mistral’s initial October 6 announcement offered an API preview and promised weights later in the month. The October 29 release makes the checkpoint inspectable and self-hostable. Mistral says further reinforcement learning since preview improved coding and defensive cybersecurity performance, while reducing the number of output tokens required on selected tasks. Those are vendor claims until a third party repeats the same tests with fixed model revisions and workloads.

The official repositories distinguish FP8 and NVFP4 checkpoints. Hugging Face also identifies GGUF conversions maintained by community and runtime teams. A GGUF conversion is not automatically interchangeable with the original checkpoint: tensor quantization, image handling, chat templates, sparse-attention implementation and tool-call parsing can all affect behavior.

The main model card identifies an image-text-to-text model. The NVFP4 card describes a Blackwell-oriented quantized checkpoint. Do not infer that a GPU supports efficient native NVFP4 simply because its software stack can load a four-bit tensor.

Why 52B active is not a 52B model

The decoder has 72 layers, a hidden width of 6,144 and 64 attention heads. The first feed-forward block is dense; the other 71 use mixture-of-experts layers. Each MoE layer has 128 routed experts, selects four for a token and also runs one shared expert. The language model thus activates a small subset of a much larger parameter pool on each token.

That design reduces compute relative to evaluating every expert, but the total weight storage remains approximately a trillion parameters. The router can choose different experts for different tokens; without a measured expert-residency policy, a system cannot assume that the remaining experts may simply be discarded.

A first-order weight-only budget illustrates the distinction:

Format Approximate 1T-parameter weight payload Important caveat
FP8, one byte per parameter about 1 TB decimal scales, mixed-precision tensors and runtime allocations add overhead
Idealized four-bit, half a byte per parameter about 500 GB decimal NVFP4 includes scale and metadata overhead; not a complete deployment estimate
52B active at FP8 about 52 GB per token’s selected weights not the total model-residency requirement

The table is dimensional arithmetic, not a measurement of the published files or a minimum GPU recommendation. In particular, NVFP4 storage is not identical to a perfectly packed four-bit array. A system may also need substantial KV-cache, temporary workspace, activation and communication buffers.

Mistral’s official Hugging Face deployment recipe uses eight H200 GPUs for an FP8 inference endpoint, and describes a 131K-token serving configuration with up to 16 concurrent requests. That concrete serving recipe is more informative than simply repeating the model’s advertised one-million-token context ceiling.

Two sparsity mechanisms, two different bottlenecks

Mistral Large 4 uses both MoE expert routing and sparse attention. The latter combines multi-head latent attention (MLA), which compresses KV representations, with a DeepSeek Sparse Attention (DSA) indexer that selects up to 2,048 key positions for a query. Hugging Face also describes shared KV state across layers and a three-to-one RoPE-to-NoPE layer pattern.

Expert sparsity reduces the feed-forward computation for each token; attention sparsity reduces the number of historical positions used in selected attention operations. Neither mechanism eliminates the need to preserve sufficient state for the intended context or to maintain a compatible cache. The exact speedup depends on prompt length, batch size, kernel support, GPU topology and the selected runtime.

For an evaluator, this produces two independent axes. A 4K-token interactive prompt may be dominated by model residency and expert dispatch. A 128K-token batch may instead expose indexer, KV-cache, prefill and cross-device communication costs. The one-million-token maximum is a supported model-context specification, not evidence of satisfactory latency at that length.

The image path is another separate budget. The release describes a 1.6B-parameter Pixtral-style vision encoder, 30 layers, 2,048 hidden width, 16 heads and 14-by-14 image patches. Image workloads should be measured separately from text-only token throughput.

Cybersecurity benchmarks: distinguish a subtest from an index

Mistral positions the model for defensive security, vulnerability reproduction and patching, coding agents and incident-response analysis. Its October 29 announcement reports 51% on the Artificial Analysis Cyber Index. The Hugging Face launch analysis separately highlights 82% on a vulnerability reproduction-and-patching subtest, as well as 93% on Cybench. These are different measures. Reporting “82% on the Cyber Index” without the subtest qualifier would mislead readers.

The same problem applies to coding. Mistral describes a 13-percentage-point improvement to 41% on Terminal-Bench 4.0 compared with the earlier preview, while the Hugging Face launch article includes additional results on other benchmark versions. An article comparing numbers must pin the exact benchmark name, version, evaluation date, tool budget and reasoning mode before drawing a trend line.

Safety claims also need separation. Mistral reports a 96.6% average refusal rate on malicious cyber requests in its chosen evaluations. That number does not establish a low false-refusal rate on authorized defensive tasks, nor does it demonstrate immunity to prompt injection. For enterprise defensive use, measure legitimate patch completion and unsafe action rates independently.

A high model score on vulnerability reproduction does not mean a patch is production-safe. Reproducing a bug, proposing a code change, passing existing tests and preserving compatibility are separate gates. An evaluation that scores only a proof of concept can overstate practical remediation value.

The license is open weight, not an unconditional MIT grant

The model is distributed under the Mistral Ecosystem Open-Weight License, described by Mistral as a modified MIT license. The key commercial condition is explicit: a business operating a Model-as-a-Service offering must obtain a separate commercial license. That restriction is not present in ordinary MIT licensing.

The distinction affects businesses that plan to sell hosted access to the model. It is not evidence that all commercial internal use is prohibited. Before deploying, read the exact license attached to the artifact, assess whether the planned service falls within its definition, and retain the license text with the model revision. A model-card label such as “other” is not a substitute for the actual terms.

Open weights do allow a team to inspect artifacts, choose its own hosting and evaluate refusal behavior independently. They do not automatically provide sovereign operations. Data residency, training provenance, runtime supply chain, network egress, logging and access controls still depend on the deployment.

A realistic self-hosted evaluation plan

Start with artifact integrity, not a chatbot screenshot. Pin the exact model revision, record file checksums and identify the checkpoint format. Verify that the selected runtime actually supports the architecture’s attention indexer, vision path, reasoning template and tool-call protocol. A successful model load is necessary but not sufficient.

Next, test a short deterministic task across the FP8 reference and each quantized variant. Record token IDs where available, answer correctness, tool-call JSON validity and image-grounding behavior. Differences may arise from quantization, sampling or runtime implementation; use the same prompts and decoding configuration before assigning causality.

For performance, separate time to first token, prefill tokens per second, decode tokens per second, peak memory and total throughput under concurrency. Test at 4K, 32K and 128K before extrapolating to the advertised 1M context. Repeat with cold and warm prefix caches. Document tensor/expert parallel layout, GPU model, interconnect and KV-cache settings.

For cyber utility, use authorized, isolated, reproducible software defects and paired benign tasks. Measure whether the model can explain the bug, reproduce it in a safe test harness, propose a minimal patch and pass regression tests. Also record false refusals and unsupported claims. Do not turn a vendor benchmark into a claim that the model was independently tested on a production environment.

Finally, assess total cost: GPU memory, interconnect, energy, operator time and capacity utilization versus the published hosted API price. A large sparse model can offer attractive active-parameter compute while remaining expensive to keep resident for low-concurrency traffic.

Decision: important release, demanding deployment

The October 29 release is a genuine Model Spotlight event because the weights and licensing changed materially after the preview. Its architecture is technically distinctive and its defensive-security emphasis deserves independent evaluation. But it is not a 52B model that conveniently fits a 52B deployment budget, and benchmark leadership should not be accepted without matching evaluation conditions.

The strongest immediate value for most readers is knowing what to test and what not to infer: sparse compute is not sparse storage; model context is not serving capacity; open weight is not unconditional open source; and vendor security scores are not proof of safe production remediation.

Sources and verification boundary

The October 29 Mistral release establishes the weight launch, pricing, license condition and current vendor claims. The October 6 preview establishes the earlier API-only stage. The official FP8 model card, NVFP4 model card and Hugging Face technical launch analysis establish architecture, available formats, examples and the distinct benchmark figures. No checkpoint was downloaded or independently benchmarked for this article.

Sources and further reading