Key takeaways
- Mistral published the improved Mistral Large 4 weights on October 29, 2026. The public ungated repository is mistralai/Mistral-Large-4-1T-A52B; official FP8 and NVFP4 variants are available, alongside community GGUF conversions.
- The model is a natively multimodal mixture of experts with about one trillion total parameters and 49 billion active parameters per token. It accepts text and images, emits text, supports a one-million-token context and exposes low, medium and high reasoning modes.
- Managed access is live through model ID mistral-large-4-0, with mistral-large-4 documented as an alias. Mistral lists $0.70 per million input tokens, $0.07 cached input and $2.10 output; those API prices do not include self-hosting infrastructure.
- Mistral reports 19.7% on private FrontierCode 2.0, 41% on Terminal-Bench 4.0, 48% on the Artificial Analysis Coding Agent Index and 51% on its Cyber Index. These are reported results, not AccessAllGPT reproductions.
- The Mistral Ecosystem Open-Weight License permits broad commercial use, but a business operating Model-as-a-Service must obtain a separate commercial license from Mistral before commercial use. It is not ordinary MIT licensing.
- The launch page says a full technical report is coming soon. Teams making high-stakes deployment decisions should wait for that report, run workload-specific evaluations and budget the substantial serving footprint rather than infer readiness from benchmark rank.
The improved weights and API are available now
On October 29, 2026, Mistral released the weights of an improved Mistral Large 4, nicknamed Le Chonk, after an earlier public preview. The canonical open-weight repository is mistralai/Mistral-Large-4-1T-A52B. At review time it was public, ungated and pinned at commit f12dabdc0c1e99e7f7924461a6f308859b05af4d, with 54 FP8 safetensor shards listed in the repository.
Managed access is also live. Mistral’s documentation names mistral-large-4-0 as the model ID and mistral-large-4 as an alias. The release points to Mistral Studio and API, OpenRouter, Microsoft Foundry and NVIDIA-hosted endpoints; self-hosters can use the Hugging Face weights with vLLM or SGLang. Availability is therefore broader than a waitlist announcement, although each provider has its own terms, regions and operational envelope.
One trillion parameters does not mean one trillion active per token
The model card describes a natively multimodal mixture-of-experts model with roughly one trillion total parameters and 49 billion active parameters per token—52 billion when embeddings and output layers are counted. Its routing design has 128 routed experts, four active per token, plus one shared expert. It accepts text and images and produces text.
Mistral documents a one-million-token context window, more than 160 languages, native function calling, JSON output and strict JSON-schema enforcement. Three reasoning modes expose an explicit cost-performance control: low for simple work, medium for everyday tasks and high by default for the hardest work. A large advertised context and sparse activation reduce neither memory planning nor the need to test long-context retrieval quality.
The API price is explicit; self-hosting cost is not
Mistral lists $0.70 per million input tokens, $0.07 per million cached-input tokens and $2.10 per million output tokens, effective with the release. Those figures establish the metered Mistral API price, not the total cost of an agent: tool calls, retries, retained context, storage and surrounding services still count.
Open weights create deployment choice rather than free inference. The reviewed sources do not publish a universal minimum GPU count, throughput target or cost per request for FP8, NVFP4 or GGUF deployment. Capacity planning must use the selected artifact, serving engine, context target, concurrency, latency objective and hardware. Compare a measured hosted bill with a measured self-hosted bill; do not treat parameter activation alone as a hardware estimate.
Coding and security gains are reported, not independently reproduced here
Mistral reports 19.7% on Cognition’s private FrontierCode 2.0, 48% on the Artificial Analysis Coding Agent Index while using 47% fewer tokens per task than GLM-5.3, and 41% on Terminal-Bench 4.0—13 percentage points above the preview with 30% fewer output tokens. It also reports a 51% Artificial Analysis Cyber Index result. These numbers are useful evaluation leads, but private data, agent scaffolds, reasoning settings and token accounting can change what they predict.
The company further reports 96.6% average refusal across cyber-related JailbreakBench, StrongREJECT and AgentHarm prompts, and 97.1% attack resistance on the B3 Agent Security Benchmark. In the same announcement, Mistral says it saw potential uplift for real attacks in live-traffic monitoring and is modifying guardrails. Builders should preserve both facts: higher reported defensive capability and explicit residual misuse risk.
The license has a material Model-as-a-Service condition
The Mistral Ecosystem Open-Weight License is a modified MIT-style license. It grants rights to use, copy, modify, distribute, sublicense, sell, train and create derivatives, subject to attribution and warranty terms. It then adds a condition: if a licensee or affiliate operates a Model-as-a-Service business, it must obtain a separate Mistral commercial license before using the model or derivatives commercially.
That condition requires legal classification before architecture selection. An internal deployment and a customer-facing hosted model endpoint may have different implications; AccessAllGPT is not providing legal advice. Record the exact license revision with the model commit, describe the offered service and obtain counsel or Mistral’s written determination before building a commercial hosted-model product.
The technical report and independent workload evidence are still missing
Mistral says the model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its European datacenters. The model card supplies architecture and benchmark notes, but the launch explicitly says a full technical report is coming soon. The reviewed materials do not yet provide a complete training-data account, independent safety audit, reproducible end-to-end cost study or universal hardware recipe.
Hugging Face documents official FP8 and NVFP4 variants and community GGUFs, plus Transformers, vLLM, SGLang and llama.cpp paths. Format availability is not deployment validation. Before production, verify artifact hashes, serving compatibility, multimodal preprocessing, tool-call behavior, context degradation, language slices, refusal behavior and license obligations on the exact revision.
What AI teams should do next
Trial Mistral Large 4 when open weights, European deployment control, multimodal input or coding-agent economics are strategically important. Pin the repository commit, select one artifact and serving route, and compare low, medium and high reasoning on representative tasks against the current model. Measure task success, severe failures, tokens, latency, concurrency, GPU memory and total cost rather than benchmark rank alone.
Constrain tool permissions and cyber workflows even if published guardrail scores look strong. Wait for the technical report when training transparency or high-assurance safety evidence is a procurement gate. Reject deployment if the license classification is unresolved, the serving footprint misses the budget, or local evaluation fails agreed quality and safety thresholds. Open weights expand control; they do not transfer the vendor’s benchmark claims into your production environment.
Copy-ready Mistral Large 4 evaluation record
Complete this record for one immutable model revision, artifact, serving route, reasoning mode and workload before approving a trial or production deployment.
Entries stay in this browser tab and are not submitted to AccessAllGPT. Blank responses are copied as [Unresolved].
Model ID and alias, Hugging Face repository, immutable commit, artifact format, files and hashes, tokenizer and chat template.
License revision, attribution, intended internal or external service, Model-as-a-Service classification, counsel or vendor determination and derivative obligations.
Mistral API, partner endpoint or self-hosted engine; region, data path, retention, quotas, support, hardware, quantization and rollback.
Text and image inputs, languages, context distribution, tools, JSON schema, autonomy, reasoning mode, latency and availability target.
Representative and adverse cases, baseline, task success, groundedness, visual and multilingual slices, long-context degradation and severe failures.
Prompt injection, tool abuse, data exfiltration, cyber-use boundaries, refusals, sandboxing, least privilege, human approval and incident response.
Input, cached-input and output tokens; retries and tools; API bill or GPU utilization; memory, throughput, concurrency, operations and break-even assumptions.
Trial, constrain, wait or reject; approved scope, owner, expiry, stop conditions, unresolved evidence and technical-report revalidation trigger.
Primary sources
Browse the publication-wide evidence index →
- Mistral Large 4 (Le Chonk) is now open weightMistral AI · Reviewed: October 29, 2026 publication metadata; weight release and availability; coding, cyber, finance, browsing and agent-security claims; reasoning modes; managed and self-hosted routes; API pricing; license summary; technical-report status · Retrieved · Supports: Mistral announced the improved Mistral Large 4 weights on October 29, described three reasoning modes and multiple deployment routes, published API pricing, and reported benchmark and safety results. The page says a full technical report is still forthcoming.
- Mistral Large 4 model documentationMistral AI Documentation · Reviewed: Model version and public-preview label; API aliases; 1M context; input, cached-input and output prices; supported modalities; structured output, function calling, document Q&A, batching, agents and built-in-tool endpoints · Retrieved · Supports: The live documentation lists Mistral Large 4 version 26.10, aliases mistral-large-4 and mistral-large-4-0, a one-million-token context window, text and image input with text output, and API prices of $0.70 input, $0.07 cached input and $2.10 output per million tokens.
- Mistral-Large-4-1T-A52B model card and repositoryMistral AI on Hugging Face · Reviewed: Public ungated repository at commit f12dabdc0c1e99e7f7924461a6f308859b05af4d; model identity; architecture and active parameters; training-compute statement; modalities; languages; 1M context; tool and JSON support; FP8 files; inference-engine support; benchmark methodology notes · Retrieved · Supports: The public ungated repository exposes 54 FP8 safetensor shards at the reviewed commit. Its model card describes a one-trillion-parameter multimodal mixture-of-experts model with 49 billion active parameters per token, 128 routed experts with four active plus one shared expert, and support for Transformers, vLLM, SGLang and llama.cpp.
- Mistral Ecosystem Open-Weight License for Mistral Large 4Mistral AI on Hugging Face · Reviewed: License text at repository commit f12dabdc0c1e99e7f7924461a6f308859b05af4d; use, copy, modification, distribution, sublicensing and sale grant; attribution condition; Model-as-a-Service commercial-license condition; warranty and liability terms · Retrieved · Supports: The modified MIT-style license permits broad use, modification, distribution and commercialization, but requires a separate Mistral commercial license when the licensee or an affiliate operates a Model-as-a-Service business and uses the model or derivatives commercially.
- Welcome Le Chonk: Mistral Large 4 on Hugging FaceHugging Face · Reviewed: October 29, 2026 publication date; FP8 and NVFP4 official variants; community GGUFs; architecture overview; benchmark discussion; Transformers, vLLM and llama.cpp support; fine-tuning and Inference Endpoints sections · Retrieved · Supports: Hugging Face independently confirms that the weights are published in official FP8 and NVFP4 variants, describes community GGUF conversions, and documents current inference and fine-tuning integration paths. It does not provide an independent benchmark reproduction.
Limitations
This is a first-party launch, documentation, artifact and license review—not independent model testing. AccessAllGPT did not download or hash the checkpoint; inspect all 54 FP8 shards or NVFP4 and GGUF variants; call the Mistral API; verify aliases through authenticated requests; run Transformers, vLLM, SGLang, llama.cpp, OpenRouter, Microsoft Foundry or NVIDIA endpoints; reproduce FrontierCode, Coding Agent Index, Terminal-Bench, Cyber Index, safety, finance, browsing or agent-security results; test multimodal, multilingual, long-context, tool-use or structured-output behavior; validate training infrastructure, data, lineage or safety controls; estimate a universal hardware requirement; obtain the forthcoming technical report; or provide legal advice on the Model-as-a-Service clause. Model files, provider availability, prices, aliases, documentation and licenses can change. All performance and safety figures are vendor- or evaluation-provider-reported unless explicitly identified as observable repository metadata.
Disclosures
AccessAllGPT received no weights, API keys, credits, hardware, managed access, private benchmark data, specialized cybersecurity-model access, commercial-license advice, briefing, review, payment or compensation for this article. Mistral AI, Hugging Face, Microsoft, NVIDIA, OpenRouter and the benchmark providers did not sponsor, review or endorse it. AccessAllGPT did not test, score or rank Mistral Large 4. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with Mistral AI or the named providers. Publication-wide relationships are listed on the disclosures page.
Further AccessAllGPT guidance
Continue the research
Get evidence-led updates for teams making production AI decisions.