Mistral AI•Rank #21NEW

Mistral Large 4

Open Weights • Mistral Ecosystemgeneralreasoningcodingmultimodalopen-source

Leaderboard Score

56.5

Scale: 0 - 100

Throughput Speed

110 tps

Context Window

API Pricing / 1M

.60

Parameter Scale

1.05T MoE / 52B

Monthly API Cost Calculator

Mistral Large 4 rates: $0.4/1M input • $1.2/1M output. Pick a workload preset or enter your own token volumes.

Monthly input (M tokens)Monthly output (M tokens)

Same workload on published-rate alternatives

Qwen3.5-397B-A17B (Qwen / Alibaba)

Qwen3.6-27B (Qwen / Alibaba)

Open the comparison engine →

Estimates use listed API rates at time of writing; providers may differ in caching discounts, rate limits, and batching pricing.

Sourced AI Radar & Benchmarks•9 Evaluated Metrics

Capability Radar & Benchmark Scores

Multi-Axis Capability Radar

Comparing Mistral Large 4 against Frontier Top 10 & Open-Source average

0 - 100 Scale

Mistral Large 4

Frontier Top 10 Avg

Open-Source Avg

Performance Vectors

Where Mistral Large 4 Leads

Coding & SWETop 10: 80% • OS: 51.1%

61.7%-18.3% vs Top 10

Reasoning & MathTop 10: 89.2% • OS: 52.7%

91.2%+2.0% vs Top 10

Agent & Tool AutomationTop 10: 78.3% • OS: 36.4%

48.0%-30.3% vs Top 10

Terminal & CLI ExecutionTop 10: 66.8% • OS: 46.2%

41.0%-25.8% vs Top 10

MMLU & Knowledge QATop 10: 50% • OS: 86.7%

70.6%+20.6% vs Top 10

Cyber & Security ExploitsTop 10: 86.2% • OS: 81.2%

82.0%-4.2% vs Top 10

All Verified Benchmark Evaluations

Direct Pass@1 and resolution rates across standardized evaluations.

OverallCodingReasoningAgenticTerminalSecurity

Featured Compute Platform

Run Benchmarks on Dedicated AI Clusters

Zero-cold-start inference endpoints with dedicated GPU instances.

Compare With Alternatives →

Executive Analysis & Verdict

Mistral Large 4 by Mistral AI is a 1-trillion parameter open-weights multimodal MoE model ranking #15 globally with a 56.5 score, 41% Terminal-Bench 4, and 82% CyberGym.

Key Strengths & Highlights

•European sovereign open-weights architecture with 1.05T MoE / 49B-52B active parameters
•Native multimodal comprehension with Pixtral vision encoder across a 1M token context
•Superior defensive cybersecurity and terminal coding execution without refusal loops

Limitations & Considerations

•High total parameter count requires multi-node cluster hardware for full FP16 hosting

Recommended Workflows

Sovereign enterprise infrastructure, defensive cybersecurity audits, and multimodal document reasoning

Architectural Overview & Capability Profile

Mistral Large 4 (codenamed “Le Chonk”) is Mistral AI’s flagship open-weights Mixture-of-Experts foundation model. Built from scratch on European sovereign AI infrastructure powered by 3,800 NVIDIA Grace Blackwell GPUs, Mistral Large 4 features 1.05 trillion total parameters with 49 to 52 billion active parameters per token. Across the RankLLMs benchmark index, Mistral Large 4 achieves an overall rating of 56.5 / 100, standing as the highest-performing open-weights foundation model developed in Europe.

Released under the Mistral Ecosystem Open-Weight License, Mistral Large 4 features native multimodal capabilities powered by an integrated 1.6B Pixtral vision encoder and an extensive 1,000,000-token context window. This allows direct native comprehension of dense technical schematics, architectural blueprints, financial reports, and multi-thousand-line repositories.

Benchmark Performance: Coding, Cyber & Independent Intelligence

Mistral Large 4 shows distinguished strength in real-world software engineering, cybersecurity, and open-weights reasoning:

Agentic Coding & Terminal Execution:
- Terminal-Bench 4.0: Reaches 41.0%, establishing leadership among non-Chinese open-weight foundation models.
- DeepSWE v1.1: Scores 61.7%, demonstrating capable multi-step file modifications and regression test validation.
- SWE-Atlas-QnA: Reaches 59.4%, effectively navigating complex legacy repositories and answering architectural dependency queries.
Defensive Cybersecurity & Security Patching:
- CyberGym: Achieves 82.0%, excelling at reproducing vulnerabilities, identifying buffer overflows, and generating verified security patches. Mistral Large 4 demonstrates virtually zero false-positive safety refusals on authorized penetration testing and audit tasks.
General Intelligence & Formal Reasoning:
- Artificial Analysis Intelligence Index v5.0: Scores 51, demonstrating competitive standing among top-tier global frontier models.
- GPQA Diamond: Scores 91.2%, showcasing strong postgraduate scientific and mathematical reasoning.
- Math500: Reaches 86.4%, solving rigorous competition-level mathematics proofs.

Sovereign Deployment & API Economics

Mistral Large 4 is distributed via open-weights on Hugging Face (with official GGUF, AWQ, and FP8 quantizations for vLLM and TensorRT-LLM) and through the Mistral La Plateforme API:

API Pricing: $0.40 per 1M input tokens and $1.20 per 1M output tokens (blended ~$0.60/1M tokens).
Self-Hosting: With 49B-52B active parameters per token, 4-bit and 8-bit quantizations can be hosted on single 8x H100/H200 nodes, delivering ~110 tokens per second with full European data sovereignty.

Frequently Asked Questions About Mistral Large 4

What is Mistral Large 4 and who developed it?↓

Mistral Large 4 is a mistral ecosystem large language model developed by Mistral AI. It ranks #21 on the RankLLMs composite leaderboard, ahead of 80% of the 95 models we track, with an overall score of 56.5 / 100.

Can Mistral Large 4 handle very long documents and codebases?↓

Mistral Large 4 supports a 1.0M-token input context window and generates about 110 tps tokens per second in our reference measurements and which is enough to hold entire repositories, long transcripts, or multi-document research sets in a single session..

Can I self-host Mistral Large 4?↓

Yes. Mistral Large 4 is released with Mistral Ecosystem licensing, so it can be self-hosted on private infrastructure with no per-token API cost. Check the specifications above for the memory footprint and quantization headroom to plan your hardware.

What is Mistral Large 4 best at?↓

Mistral Large 4 is strongest in European sovereign open-weights architecture with 1.05T MoE / 49B-52B active parameters; Native multimodal comprehension with Pixtral vision encoder across a 1M token context; Superior defensive cybersecurity and terminal coding execution without refusal loops. Full pillar-by-pillar numbers are in the benchmark tables above.

What should you watch out for with Mistral Large 4?↓

Known trade-offs include: High total parameter count requires multi-node cluster hardware for full FP16 hosting. Weigh these against the strengths above for your specific workload.