Mistral Large 4 (Le Chonk): Price, Free Access and How to Run the 1T Open Model
Mistral Large 4, nicknamed "Le Chonk", is a 1-trillion-parameter open-weight model whose weights went up on Hugging Face on October 29, 2026. The weights are free to download, but running them takes datacenter GPUs or at least 256 GB of memory for the smallest 1-bit build. For most people the practical route is the paid Mistral API (model ID mistral-large-4) or OpenRouter at $0.70 per 1M input tokens and $2.10 per 1M output tokens. aiwith does not offer Mistral Large 4, so an aiwith subscription will not unlock it.
Last checked: October 30, 2026. This is an access and pricing guide built from the official Hugging Face model card, Unsloth's GGUF page and OpenRouter's live model listing. It is not a hands-on review.
What is Mistral Large 4?
Mistral Large 4 is Mistral AI's largest model to date. According to the official model card:
- Size: about 1 trillion total parameters in a Mixture-of-Experts design, with 49 billion active parameters per token (52 billion including embeddings and output layers). It uses 128 routed experts, 4 active per token, plus 1 shared expert.
- Context: 1M tokens.
- Input and output: text and image in, text out. The vision encoder handles variable image sizes and aspect ratios.
- Reasoning: one set of weights for chat, reasoning, coding and agents, with reasoning effort set per request to
low,mediumorhigh(the default). - Tools: native function calling, JSON output and strict JSON schema enforcement.
- Languages: Mistral says it is fluent in 160+ languages, including every official EU language.
- Training: trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters.
Mistral says the model is especially strong on cybersecurity, finance and law, and claims it beats frontier closed models on visual grounding. Those are vendor claims; the model card links to Mistral's blog for the supporting benchmarks.
Is Mistral Large 4 free?
The weights are free. Using it is usually not. Here are the options:
| Route | Cost | Who it suits |
|---|---|---|
| Download weights from Hugging Face | Free to download; you pay for the hardware | Teams with GPU servers or very large-memory machines |
| Le Chonk Gallery demo Spaces | Free, with a small shared budget per visitor during the release's first week | Anyone who wants a quick test |
Mistral API (mistral-large-4) | Paid per token | Developers who want Mistral's own hosting |
OpenRouter (mistralai/mistral-large-4-0) | $0.70 input / $2.10 output per 1M tokens | Developers already using OpenRouter |
The free demo budget is the only zero-cost way to try the model without your own hardware that we could confirm. Mistral describes it as temporary, covering the first week after release.
Mistral Large 4 pricing
OpenRouter's live listing on October 30 shows:
| Per 1M tokens | |
|---|---|
| Input | $0.70 |
| Cached input | $0.07 |
| Output | $2.10 |
A simple mix of 1M input tokens and 1M output tokens costs about $2.80. Cached input is 90% cheaper than fresh input, which matters for agents that resend long transcripts and tool results.
Some earlier coverage quoted a higher list price of $1.36 input and $4.18 output per 1M tokens from the October preview period. The current OpenRouter rate matches the cut reported alongside the open-weight release. Prices can change quickly after a launch, so confirm the rate on Mistral's own pricing page before you build a budget.
The weights come under the Mistral Ecosystem Open-Weight License, a modified MIT license. It allows commercial and non-commercial use, but running a Model-as-a-Service business needs a separate commercial license from Mistral. Read the license file before you resell access to it.
How do I use Mistral Large 4?
1. Mistral API
The model card says the model is served on the Mistral API as mistral-large-4. Mistral recommends temperature=1.0 and top_p=0.95 at every reasoning effort: low for fast chat, medium for daily use and high for complex prompts and agents.
2. OpenRouter
Use the model ID mistralai/mistral-large-4-0. OpenRouter lists a 1,048,576-token context and text plus image input.
curl https://openrouter.ai/api/v1/chat/completions \
-H "Authorization: Bearer $OPENROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "mistralai/mistral-large-4-0",
"messages": [{"role": "user", "content": "Summarize this contract clause."}]
}'
3. Mistral Vibe (coding agent)
Mistral's open-source CLI coding assistant, Mistral Vibe, can run on Mistral Large 4. Install it with pip install mistral-vibe, run vibe in your project, enter a Mistral API key and pick mistral-large-4. It can also point at your own vLLM server.
4. Self-host the full model
The official examples serve the FP8 checkpoint with vLLM or SGLang across 8 GPUs using tensor parallelism. Mistral also publishes an NVFP4 checkpoint for Blackwell GPUs with a smaller memory footprint. The checkpoint includes a built-in speculative decoding head (DSpark), which Mistral says gives up to 1.9x higher throughput. Mistral has not published a minimum GPU list on the model card.
5. Run it locally with GGUF
Unsloth published GGUF builds for llama.cpp. Its memory guidance:
| Quantization | File size | Memory needed |
|---|---|---|
| 1-bit | 233 GB | 256 GB RAM device |
| 2-bit | 302-378 GB | 384-512 GB |
| Q8 (near full precision) | 1.1 TB | 1.2 TB+ |
Even the 1-bit build is far beyond a normal laptop or gaming PC. It is realistic on a large-memory workstation or server, not on consumer hardware. The GGUF repo includes the vision projector and the speculative decoding draft model.
How does Mistral Large 4 compare?
Mistral's model card compares it only with other open-weight models, using Artificial Analysis indexes:
- Artificial Analysis Coding Agent Index: 48%, which Mistral calls effectively tied with the leading open-weight model, GLM-5.3, while using 47% fewer tokens per task on average.
- Artificial Analysis Intelligence Index v5: 26, the highest among Western open-weight models. Mistral notes that the leading Chinese open-weight models score higher on this aggregate index.
The model card does not include head-to-head numbers against GPT-6 Astra, Claude, Gemini 4 Carbon or DeepSeek V4.1, so we have not invented any. In practice:
- Choose Mistral Large 4 if you need open weights you can host yourself, EU-based training and infrastructure, strong multilingual coverage, or a model you can fine-tune.
- Choose a closed frontier model such as GPT-6 Astra or Claude if you just want the strongest hosted assistant and do not need the weights.
- Choose DeepSeek V4.1 or Kimi K3 if cost per task matters most and you are comfortable with Chinese open-weight models.
Can I use Mistral Large 4 on aiwith?
No. aiwith does not include Mistral Large 4 in its model selector, and Router does not list it either. If you came here because you want a strong model in the browser today, aiwith already carries these alternatives:
- GPT-6 on aiwith: GPT-6 Astra, Sol and Luna for general work, writing and analysis.
- Claude on aiwith: Claude Fable, Opus, Sonnet and Haiku for coding and long documents.
- DeepSeek on aiwith: the value option for coding and everyday tasks.
- Kimi K3 on aiwith: a 1M-token context model for long files and agent-style tasks.
None of these is Mistral Large 4 under another name. If you need Mistral specifically, use the Mistral API, OpenRouter or the free demo Spaces above.
FAQ
Is Mistral Large 4 open source? It is open-weight under the Mistral Ecosystem Open-Weight License, a modified MIT license. Commercial use is allowed, but offering it as a hosted API service requires a separate license from Mistral.
What does "Le Chonk" mean? It is the nickname Mistral uses for Mistral Large 4 on its model card and demo gallery, a joke about the model's size.
How much RAM do I need to run Mistral Large 4 locally? Unsloth's smallest 1-bit GGUF is 233 GB and needs a device with about 256 GB of RAM. Near-full-precision Q8 needs 1.2 TB or more.
Does Mistral Large 4 support images? Yes. It accepts text and image input and returns text.
What is the context window? 1M tokens on the model card; OpenRouter lists 1,048,576 tokens.
Can I use Mistral Large 4 in Le Chat or Mistral's consumer apps? We could not confirm consumer-app availability from the sources we checked. The confirmed routes are the Mistral API, OpenRouter, Mistral Vibe with an API key, the Hugging Face demo Spaces and self-hosting.