Mistral Large 4 Launch: $1.36 MoE, Specs & Open Weights

On October 6, 2026, Paris-based Mistral AI officially launched Mistral Large 4 (internally codenamed “le Chonk”) into public preview via Mistral Studio. The release marks a major escalation in open-weight frontier computing: a 1.05-trillion-parameter Mixture-of-Experts (MoE) architecture operating with 52 billion active parameters (49 billion per token), native multimodal vision processing, and a 1,000,000-token context window.

Crucially for developers tracking compute budgets, Mistral introduced the preview API at aggressive pricing: $1.36 per million input tokens and $4.18 per million output tokens, paired with $0.14 cached input reads. Mistral committed to publishing the model weights by the end of October 2026, positioning Large 4 as the largest open-weights model in production history.

1. Quick Facts and Specifications (October 2026)

Specification	Official Metric	Technical Detail
Model Name	Mistral Large 4	Codename “le Chonk”
Release Date	October 6, 2026	Public preview on Mistral Studio
Total Parameters	1.05 Trillion	Granular Mixture-of-Experts (MoE)
Active Parameters	52 Billion	49B active per token, remainder in embeddings/output
Vision Encoder	1.6 Billion parameters	Natively multimodal (high-res image ingestion)
Context Window	1,000,000 tokens	Full 1M context supported in preview
Input Pricing	$1.36 / 1M tokens	Standard prompt ingestion
Output Pricing	$4.18 / 1M tokens	Standard completion output
Prompt Cache Read	$0.14 / 1M tokens	Automated prefix cache hits
Open Weights Status	Promised late October 2026	Available for self-hosted enterprise clusters

2. Pricing Comparison: How Mistral Undercuts Frontier APIs

The defining dynamic of the late-2026 AI market is the cost delta between proprietary frontier models and high-parameter MoE architectures. At $1.36 input and $4.18 output, Mistral Large 4 prices well below proprietary flagships:

Model SKU	Input / 1M Tokens	Output / 1M Tokens	Cache Reads / 1M	Open Weights?
Mistral Large 4	$1.36	$4.18	$0.14	Yes (End Oct 2026)
GPT-5.6 Sol (Promo)	$4.00	$20.00	$0.40	No
Claude Opus 5	$5.00	$25.00	$0.50	No
Claude Fable 5.1	$10.00	$50.00	$0.25	No
GPT-6 Astra	$10.00	$50.00	Variable	No

For agent developers running high-volume repository scanning, Mistral’s $0.14 cache read rate undercuts even Claude Fable 5.1’s $0.25 cache reads and runs at roughly one-tenth the cost of GPT-6 Astra’s $50 output pricing.

3. Architecture and Training Infrastructure

Mistral Large 4 was trained from scratch across Mistral’s European data center clusters utilizing 3,800 NVIDIA Grace Blackwell GPUs, adhering strictly to European Union sovereign data protection standards.

Key Architectural Highlights:

Granular Expert Routing: Unlike earlier coarse MoE models that routed tokens to 2 out of 8 experts, Mistral Large 4 utilizes a fine-grained gating network that dynamically selects smaller sub-expert blocks, maintaining 49B active tokens out of 1.05T total parameters.
Integrated Vision Backbone: The 1.6B parameter vision encoder is embedded directly into the intermediate projection layers rather than operating as a detached wrapper, allowing native document parsing, UI screenshot navigation, and diagram analysis.
1M Context Needle-in-a-Haystack: Mistral reports >99.4% retrieval accuracy across the full 1,000,000 token context window, preventing context degradation during multi-document synthesis.

4. The Open-Weights Commitment and Enterprise Impact

The critical differentiator for enterprise teams is the promised open-weights release scheduled for late October 2026:

Self-Hosting Feasibility: While 1.05T total parameters cannot run on consumer hardware, a 52B active parameter footprint means enterprise clusters running 8x or 16x 80GB/192GB GPUs can serve Mistral Large 4 with FP8/INT4 quantization.
Air-Gapped Compliance: Highly regulated industries (defense, banking, healthcare) that cannot pipe proprietary data through US-hosted APIs like OpenAI or Anthropic gain an open-weights alternative with frontier-grade reasoning.
Geopolitical Positioning: As analyzed in our breakdown of China’s AI Blitz, European sovereign AI initiatives rely on Mistral providing a credible non-US alternative to proprietary API monopolies.

5. Strategic Recommendations for Engineering Teams

High-Volume Data Processing: If you currently run batch classification, bulk document extraction, or RAG pipelines on proprietary frontier models, test Mistral Large 4 in Mistral Studio today. The $1.36 / $4.18 price point offers an immediate 60-80% margin improvement.
Agentic Coding Workloads: For complex multi-file refactoring, benchmark Mistral Large 4 against Claude Fable 5.1 on your team’s specific test harness. While Fable holds top marks on Terminal-Bench, Mistral’s $0.14 cache read drastically reduces iteration costs.
On-Premise Infrastructure Planning: Teams intending to self-host should begin preparing GPU cluster configurations ahead of the late October weights drop.

Published: October 6, 2026. Verified against official Mistral AI technical disclosures, Mistral Studio API documentation, and European compute registry filings.

Primary Sources:

Mistral AI: “Announcing Mistral Large 4: Frontier Capabilities, Open Horizons” (October 6, 2026)
Mistral Studio API Reference & Pricing Card: console.mistral.ai
NVIDIA Blackwell Compute Deployment Notices: nvidia.com/enterprise