Mistral Large 4 (Le Chonk)

Mistral Large 4 (Le Chonk) is Mistral's flagship open-weight multimodal model, released on October 29, 2026. Built on a massive mixture-of-experts (MoE) architecture totaling approximately 1,000 billion (1T) parameters, the model activates just 49 billion parameters per token. This design makes Le Chonk a landmark local AI model 1000B parameters 2026 release, aimed squarely at developers who demand sovereign, self-hosted frontier intelligence without reliance on opaque cloud APIs.

The model is distributed under the Mistral Ecosystem Open-Weight License, an MIT-derived license that allows extensive self-hosting and internal enterprise modification while placing targeted restrictions on commercial model-as-a-service providers. Le Chonk combines native text and image processing with a 1,000,000-token context window, optimized specifically for autonomous software engineering, deep vulnerability research, and complex agentic workflows.

In the open-weight hierarchy, Mistral Large 4 occupies the absolute top tier for agentic capability. By combining sparse token routing with selective context attention, it challenges the largest proprietary models from OpenAI and Anthropic while giving engineers full control over deployment, data retention, and runtime parameters.

Mistral Large 4 employs a sparsely gated Mixture of Experts architecture. While the entire checkpoint comprises 1000B parameters, the routing mechanism routes each token through specialized sub-networks, activating only 49B parameters during forward passes.

This routing has a direct impact on hardware behavior:

• Memory Footprint: Governed by the full 1000B parameter set. All expert weights must reside in VRAM or system memory to avoid catastrophic offloading latency.
• Compute Overhead: Governed by the 49B active parameters. Once the weights for active experts are loaded into cache, the arithmetic throughput required per token resembles a mid-sized model rather than a 1T monolith. This decoupling is the core driver of Mistral Large 4 (Le Chonk) MoE efficiency.

Mistral pairs this expert sparsity with DeepSeek Sparse Attention (DSA) and an integrated Pixtral vision encoder. Rather than performing dense attention across the entire 1,000,000-token context window, the model uses an attention indexer to identify and attend to relevant prior positions. This dual-sparsity approach (sparse feed-forward networks via MoE alongside sparse temporal attention) prevents memory-bandwidth saturation from crippling throughput at extreme context depths.

The model also integrates multiple-teacher distillation to provide three distinct runtime reasoning modes: low, medium, and high effort. Practitioners can configure this parameter at inference time to trade token generation latency against analytical rigor.

Le Chonk is engineered for complex, multi-step agent environments rather than basic conversational chat. Its core training distribution emphasizes environments where models must inspect output, adjust assumptions, and execute code over long horizons.

Standard frontier models often refuse defensive security tasks, such as reproducing vulnerabilities or analyzing exploit chains, due to generic safety filters. Mistral Large 4 excels here, scoring 50.9% on the Artificial Analysis Cyber Index (and reaching up to 82% in vulnerability reproduction and remediation evaluations). It accepts raw binaries, network capture logs, or source trees, analyzes the exploit surface, and generates patches directly within local environments.

Practitioners targeting Mistral Large 4 (Le Chonk) for coding gain a model tuned on verifiable execution loops. Le Chonk parses terminal outputs, diagnoses build failures, navigates massive multi-file repositories, and authors clean pull requests. Its 1M token context window allows an agent to ingest an entire framework codebase, relevant documentation, and historical git diffs simultaneously.

The native Pixtral vision encoder enables Le Chonk to inspect architecture diagrams, UI design mockups, schematics, and terminal screenshots. Combined with its coding capabilities, an engineer can feed a system diagram alongside log traces to locate failure points in distributed systems.

Le Chonk maintains native fluency across major European and Asian languages, handling cross-lingual legal document parsing, technical translations, and international regulatory compliance without syntax degradation.

Deploying a 1000B parameter model on premises demands serious planning around memory bandwidth and bus topology. Understanding Mistral Large 4 (Le Chonk) VRAM requirements is essential before allocating compute.

To run Mistral Large 4 (Le Chonk) locally, plan for the following memory footprints:

• FP8 (Official Weights): ~1,000 GB VRAM. Requires an enterprise cluster such as 8x NVIDIA H100/H200 (80GB/141GB) or multi-node configurations.
• NVFP4 (Official Calibrated): ~606 GB VRAM. Calibrated directly by Mistral to match FP8 output quality. This checkpoint fits comfortably on four NVIDIA GB300 GPUs or an 8x H100 80GB node.
• GGUF Q4_K_M (Community / Unsloth): ~560-620 GB total memory. The best quantization for Mistral Large 4 (Le Chonk) for workstations combining high unified memory or multi-GPU rigs.
• GGUF Q3_K_M: ~430-480 GB total memory. Usable when memory limits require a slight compromise in reasoning accuracy.

Engineers frequently ask how to run 1000B model on consumer GPU hardware. The short answer: you cannot run it on a single consumer GPU, nor on a standard dual-RTX 4090 desktop. Even at 2-bit quantization, a 1T model will not fit into 48 GB of VRAM.

To deploy Le Chonk without a hyperscale cloud budget, consider these platforms:

• Enterprise GPU Nodes: The best GPU for Mistral Large 4 (Le Chonk) deployment is the NVIDIA H200 (141GB VRAM) or B200/GB300 series. A node with 8x H100 (80GB) provides 640 GB of high-bandwidth VRAM, enough to run the official NVFP4 release at full context.
• Prosumer Multi-GPU Workstations: Rigs equipped with 7x or 8x RTX 6000 Ada (48GB each) supply up to 384 GB of VRAM, which requires Q2 or aggressive layer offloading.
• Unified Memory Mac Clusters: Apple Silicon Macs with unified memory (such as M2, M3, or M4 Max chips) max out between 128 GB and 192 GB per machine, meaning a single Mac cannot load the checkpoint. However, networked clusters utilizing distributed inference frameworks (such as Exo or llama.cpp over fast interconnects) can distribute the GGUF across multiple high-spec Studio units.
• High-RAM Dual-Socket EPYC/Xeon Workstations: Systems containing 768 GB to 1 TB of DDR5 system RAM can run Q4_K_M via CPU offloading, though throughput will be heavily memory-bandwidth constrained.

Mistral Large 4 (Le Chonk) tokens per second performance depends on where the active experts reside:

• 8x H100 / H200 (NVFP4 via vLLM): Expect 20-35 tokens per second. Because active compute is only 49B parameters, generation speed is rapid once the weights are memory-resident across NVLink.
• Distributed Enterprise Nodes (FP8): Expect 15-25 tokens per second.
• CPU / System RAM Hybrid Offload (GGUF): Expect 0.5-2.5 tokens per second. Useful for offline, non-interactive agent evaluations, but slow for interactive chat.

For enterprise inference, vLLM provides native support for the official FP8 and NVFP4 checkpoints using tensor parallelism across 8 GPUs:

1vllm serve mistralai/Mistral-Large-4-NVFP4 \
2  --tensor-parallel-size 8 \
3  --max-model-len 32768

For quantized prototyping on workstation clusters or unified setups, use llama.cpp or Ollama with the Unsloth GGUF release:

1ollama run mistral-large-4:q4_k_m

Ensure your host system allocates at least 650 GB of swap alongside physical memory to handle model initialization.

Evaluating Mistral Large 4 (Le Chonk) vs competitor flagships highlights distinct trade-offs between sparse architectures, dense clusters, and alignment policies.

Llama 3.1 405B is a dense model. Every token generated must compute all 405B parameters, demanding continuous raw arithmetic horsepower. Le Chonk holds more total parameters (1000B), requiring similar or greater VRAM capacity to store weights, but its 49B active routing makes generation cycles substantially faster. If your infrastructure supports 600-800 GB of VRAM, Le Chonk delivers higher throughput per token and integrates native vision, which Llama 3.1 405B lacks.

DeepSeek-V3 features an MoE architecture with 671B total and 37B active parameters. While DeepSeek-V3 targets general reasoning and cost-effective serving, Le Chonk pushes context length much further (1,000,000 tokens versus 128,000 tokens) and offers significantly better utility in offensive/defensive cybersecurity. On the Mistral Large 4 (Le Chonk) reasoning benchmark evaluations for software engineering (such as Terminal-Bench and FrontierCode), Le Chonk holds a marked advantage for local autonomous agents, while DeepSeek-V3 remains easier to fit onto slightly smaller multi-GPU nodes.