Mistral Large 4
Mistral Large 4 ("ML4", internally nicknamed "le Chonk") is Mistral AI's 1.05T-parameter natively multimodal sparse MoE flagship, released as a public preview on October 6, 2026, with 52B active parameters, a 1.6B vision encoder, a 1M-token context window, up to 256K output tokens, and a rare strength in cybersecurity/agentic-automation evaluations.
Empirical Evaluation Results
Architectural Profile & Capabilities
Mistral Large 4 is Mistral AI's largest and most capable model to date, launched as a public-preview API model on 6 October 2026 under the identifiers `mistral-large-4` / `mistralai/mistral-large-4-0`; open weights were announced but had not yet shipped at the time of writing. Architecturally it is a sparsely activated, granular Mixture-of-Experts transformer with roughly 52B active parameters out of ~1.05T total (some Mistral surfaces list 49B active, an unresolved internal inconsistency), paired with a ~1.6B-parameter vision encoder for natively multimodal text+image input. It supports a 1,048,576-token context window with up to 256K output tokens, is documented as fluent in 160+ languages, and exposes tool calling and structured/JSON outputs. Mistral positions ML4 as a unified instruction + reasoning + agentic model rather than a separate reasoning SKU, and does not publish a fixed test-time-compute token budget or a named reasoning-effort dial. Independent measurement places it mid-pack on aggregate intelligence — 38 / 38.4 on the Artificial Analysis Intelligence Index preview (comparable to GPT-6 Luna max and DeepSeek V4.1 Flash max) — while it performs disproportionately well on security and business-workflow automation (Artificial Analysis Cyber Index co-first, ~59.9% AutomationBench). Long-context recall at ~81% on AA-LCR is strong relative to peers, and code-agentic numbers are respectable but clearly below frontier closed models: 61.7% DeepSWE v1.1, 59.4% SWE-Atlas-QnA, and 28.3% on the hard Terminal-Bench 4.0 suite. Legal-agentic performance (Harvey's Legal Agent 15.8%) is a notable weakness for regulated-domain deployment without heavy scaffolding. Operationally it is a value-oriented frontier-class open-weight-bound model: list pricing is $1.36/$4.18 per 1M input/output tokens (launch discount pricing of roughly $0.68/$2.09 was reported for the preview window), throughput is moderate at ~68 tokens/second on Mistral's own API, and the 1M window carries substantial latency and cost implications. Note that Mistral's scoring comes from a mix of its own harness and third-party evaluators (Artificial Analysis, Vals AI), and several standard academic suites (MMLU-Pro, GPQA Diamond, AIME, LiveCodeBench, SWE-bench Verified, MMMU) were not published for this exact variant at release.
Granular MoE: ~52B active parameters of ~1.05T total (Mistral's model card; a secondary Mistral listing cites 49B active and ~1T total, an unreconciled discrepancy). Dedicated ~1.6B-parameter vision encoder provides native image input alongside text. 1,048,576-token context window with up to 256K tokens of output. Mistral does not publicly document the attention variant (MLA vs. GQA vs. hybrid sliding-window), expert count, experts-per-token, router/aux-loss scheme, or the specific mechanism used to extend to 1M tokens. No published fixed test-time reasoning-token budget or reasoning-effort parameter; Mistral markets it as a single hybrid model unifying instruct, reasoning, and agentic behavior. Supports tool/function calling, structured JSON outputs, and 160+ languages.
Recommended Workloads & Primary Use Cases
- •Terminal/agentic ceiling is low for a 1T-class model: ~28.3% on Terminal-Bench 4.0 (Mistral-reported; 26.8% under Artificial Analysis' own harness). Do not assume it can complete long-horizon autonomous terminal or CI/CD repair loops without a strong external scaffold.
- •Long-context prefill is the dominant cost/TTFT driver. At 1M tokens, per-call cost reaches ~$1.36 before any output, and first-token latency rises materially; the effective context is better budgeted at a few hundred thousand tokens. Mistral has not published an MRCR-style needle-in-haystack score for this variant, so deep-context fidelity beyond ~100K tokens is unverified.
- •It is a public preview with no published benchmark coverage for MMLU-Pro, GPQA Diamond, AIME 2025/2026, LiveCodeBench, SWE-bench Verified, MMMU, or DocVQA. Treat any such numbers circulating online as family proxies or misattributions — several aggregator pages conflate Large 4 with larger/other Mistral checkpoints and with unrelated Terminal-Bench 2.1 figures.
- •Reasoning behavior is not explicitly controllable: there is no documented reasoning-effort dial, no separate reasoning-token billing field, and the format of any hidden reasoning traces is unspecified. Prompt/parsing logic built for GPT-6 or Claude Opus 5 reasoning blocks will not transfer directly.
- •Weak regulated-domain agentic accuracy: 15.8% on Harvey's Legal Agent Benchmark and 31.7% on Vals AI Legal Research Bench mean legal/financial agent workflows need retrieval grounding, human-in-the-loop review, and per-step validation rather than end-to-end delegation.
- •Licensing and weight availability were unresolved at launch: presented as open-weight-bound, but weights had not been released and the license terms were unpublished, so on-prem/weight-based capacity planning is premature.
Calls to Mistral Large 4 are routed through strict Zero Data Retention inference channels. Prompts and outputs are never stored, indexed, or monitored by Mistral AI.
More Models from Mistral AI
Compare other engines available in this laboratory.