Mistral Large 4 Open Weights Released: 1T Params, API Halved, 47% Fewer Coding Tokens — How to Choose?

TL;DR

Mistral AI released open weights for Large 4 (codename "Le Chonk") on October 29. It's a trillion-parameter open-weight MoE flagship: ~1T total params, 52B active, 1M token context window, natively multimodal (text + image input), covering 160+ languages. API pricing dropped from ~$1.40/$4.20 to $0.70/$2.10 per million tokens (input/output) — nearly halved. Trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in European data centers. Modified MIT license: free to use and commercialize, but MaaS resale requires separate commercial license.

Parameters & Architecture: Europe's Trillion-Parameter Flagship

Core specifications:

Training was entirely on European infrastructure, from scratch — not a fine-tune of an existing open model. The 3,800 Grace Blackwell scale places it in the top tier of open-weight training investment. After pre-training, the model underwent approximately three weeks of reinforcement learning (RL), with Mistral reporting that gains were still accelerating with no sign of saturation.

Benchmarks: Coding Efficiency & Cybersecurity Dual Narrative

Coding Performance

Token efficiency is the standout: achieving the same Coding Agent Index (48%) as the leading open-weight model while consuming 47% fewer tokens per task. Terminal-Bench 4.0 also reduced output tokens by 30%. At API pricing, this means the actual cost per coding task is less than half of competitors.

Cybersecurity

Cyber Index measures the model's ability to discover vulnerabilities in codebases, reproduce them, and patch them without breaking existing functionality. 51% is leading among open-weight models. The 96.6% jailbreak rejection rate improved from 95% in the preview version. Mistral also provided preview access to specific variants for cybersecurity partners like Orange Cyberdefense, for legitimate penetration testing scenarios under strict contractual constraints.

Enterprise Benchmarks

API Pricing & Deployment: Cloud to Local

API Pricing (per million tokens)

Compared to the preview version, this is nearly a 50% reduction. Compared to closed-source flagships at similar capability (GPT-6 Astra output ~$10/M tokens), the output price is approximately one-fifth.

Deployment Channels

The model also provides NVFP4 quantized variants for higher-throughput deployment scenarios. Self-hosting requires sufficient GPU memory to load the trillion-parameter weights — while MoE means only 52B activate during inference, all weights must be loaded into memory.

License: Modified MIT with MaaS Restriction

Mistral Large 4 uses the Mistral Ecosystem Open-Weight license (modified MIT):

• ✅ Free to use, modify, fine-tune
• ✅ Commercial deployment (integrate into your own product)
• ✅ Enterprise intranet deployment
• ❌ Reselling as "Model as a Service" (MaaS) requires separate Mistral commercial license

If you integrate the model into your SaaS product (where the model is part of the product, not the product itself), you're fine. But if you want to run a "Mistral Large 4 API relay service" selling direct API access, you need a commercial license.

This restriction is Mistral's commercial moat: open-source ecosystem builds community and developer adoption for free, but cloud providers wanting to resell the API must pay. Compared to Llama 4's Apache 2.0 (no such restriction), Mistral chose a more conservative commercial strategy.

Ecosystem Endorsements

Mistral Large 4 received backing from cybersecurity and sovereign cloud giants:

• Booz Allen (US government cybersecurity contractor)
• Cisco (network infrastructure)
• Cloudflare (edge cloud and security)
• CrowdStrike (endpoint security)
• Elastic (search and observability)
• NVIDIA (GPU and inference infrastructure)

These partners share a common need: data residency, compliance auditing, or intranet deployment. Open weights mean the model can run in your own data center under your own policies, without sending sensitive code and security logs to external APIs.

Competitive Comparison: Open-Weight Flagship Landscape

Mistral Large 4's differentiation: trillion-parameter open weights + coding token efficiency + specialized cybersecurity capabilities + European sovereignty. These four dimensions have almost no direct competitor.

Developer Action Guide

1. Try immediately: Test on Mistral Studio or OpenRouter for free, compare token consumption and output quality on coding tasks
2. Evaluate self-hosting feasibility: Trillion-parameter weights require significant VRAM; test NVFP4 quantized variants on multi-GPU A100/H100 setups
3. Validate cybersecurity use cases: If your team does penetration testing, benchmark the model's vulnerability discovery and patching capabilities using Cyber Index-related benchmarks
4. Review license terms: If your product involves providing API services externally, confirm whether MaaS terms apply and contact Mistral for commercial licensing if needed
5. Prepare for fine-tuning: After setting up inference with vLLM or SGLang, run LoRA fine-tuning on domain-specific data to test vertical scenario gains

Three Points of Caution

1. Benchmarks are self-reported: The numbers above come primarily from Mistral's official channels and partner scenarios. Independent community reproduction takes time. Third-party indices have previously scored open-weight flagships more conservatively.
2. Trillion-parameter MoE ≠ everyone can run it locally at full strength: Consumer-grade quantization lowers the barrier, but enterprise-grade inference costs, operations, and security hardening remain real thresholds. "Downloadable" and "runnable, manageable" are different things.
3. Open source is a double-edged sword: The model is resolute in rejecting black-box attacks, but Mistral acknowledges potential offensive capability increases and is adjusting guardrails. Specialized variants for penetration testing scenarios remain under strict review.

Industry Significance

Mistral Large 4's release sends signals on three levels:

Technical: Open-weight models have officially entered the trillion-parameter era. The previous largest open-weight model was Llama 4 at 400B; Large 4 pushes this ceiling to 1T. MoE architecture makes trillion-parameter inference cost approach that of a 52B dense model.

Commercial: Mistral's "open weights + MaaS restriction" strategy preserves open-source ecosystem vitality while protecting its commercial moat. This contrasts with Llama 4's fully permissive Apache 2.0 — two open-source business models placed side by side in the same week.

Geopolitical: Trained from scratch in European data centers, covering 160+ languages, backed by NATO-country cybersecurity giants — this is substantive progress for European sovereign AI. For banks, public sector, and regulated industries, this isn't just marketing language; it's a hard requirement on procurement and risk management checklists.

Similar to Aleph Alpha Kolibri's sovereign AI route, Mistral Large 4 further cements "European sovereign AI" as an independent track. But their paths differ: Kolibri pursues mid-scale (78B) + Merlin-Arthur hallucination protocol refinement, while Large 4 pursues trillion-parameter + full-stack open scale.