Mistral Large 4 open weights: providers, pricing, benchmarks

Mistral has released an improved Mistral Large 4 checkpoint as open weights. As of November 3, 2026, Mistral's API lists the model at $0.70 per million fresh input tokens and $2.10 per million output tokens; our published coding benchmark reports 78% of tasks solved for Large 4 and 81% for Kimi K3.1 Pro. The Large 4 benchmark checkpoint (preview or released weights), quantization and resolved provider remain unconfirmed (Mistral pricing, Requesty benchmark).

Mistral opened the public preview on October 6, 2026, with a two-week discount, and later published the improved checkpoint (changelog, weight release). The prices and results below need two caveats: a gateway can list Large 4 without hosting inference itself, and a preview benchmark does not measure the released checkpoint.

What is Mistral Large 4?

Mistral Large 4 is a multimodal model with 1.05 trillion total parameters and 52 billion active parameters. Its model card also lists a 1.6 billion parameter vision encoder and a 1 million token context window.

The model supports function calling (the model can request a tool or function), structured outputs, chat completions and batching. According to Mistral, the released checkpoint improves on the preview in coding, defensive cybersecurity and reasoning efficiency.

Hugging Face lists two official weight variants, FP8 and NVFP4 (Hugging Face). Mistral recommends vLLM or SGLang for deploying the downloaded weights, which you can do without Requesty.

Where can you access Large 4?

Start with the API route, then check who serves the tokens. The listings were checked on November 3, 2026, and all token prices are in US dollars per million tokens; OpenRouter's provider and pricing listing is the source for its two rows.

The Requesty figures are catalog-listed upstream rates, not confirmed billable rates. They match the rates in an earlier Mistral model card snapshot, while Mistral's current pricing shows $0.70 and $2.10. That match does not explain the catalog difference or establish what Requesty bills now, so confirm the rate with Requesty billing or against a current invoice.

For the listed endpoints, OpenRouter and Requesty route to Mistral; neither listing makes them a separate inference host. NVIDIA is a separate host with a prototype API, although its per-token prices and hosted quantization are unconfirmed, so there is no confirmed production price spread between independent hosts for this comparison. We have examined the provider-variance problem that arises when the same weights are served across providers.

OpenRouter's EU quote costs more than its standard listing and requires a Business or Enterprise plan. Requesty lists mistral-large-4@eu as an EU-only ID; verify the endpoint and your contract before making a residency claim for a deployment.

How does Large 4 compare with Kimi and GLM?

Our coding benchmark tested 25 models on 129 backend tasks. Every model used the same coding agent, tools and 80-turn budget.

Kimi K3.1 Pro leads Large 4 by 3 percentage points on the published scores. The summary gives GLM-5.3 Flash a median turn time, but that figure cannot tell us how many tasks GLM solved or how its solve rate compares with Large 4.

There is a bigger caveat for Large 4: the summary does not say whether the run used the preview or the released checkpoint. Its quantization and resolved provider are also unconfirmed, so we cannot attribute the 78% result to the released weights.

Mistral separately published a blind human coding-quality evaluation for the preview. It uses a different test and scoring scale, and its Kimi K3 is a different model from the Kimi K3.1 Pro in our table.

Source: Mistral's preview announcement.

For the released weights, Mistral reports a 19.7% score on Cognition's private FrontierCode 2.0 benchmark. It also reports that the released model resisted 97.1% of attacks on the B3 Agent Security Benchmark, compared with 93.3% for the preview (Mistral's weight release). The B3 figure is a security result and cannot be compared with a coding-task solve rate.

What would a workload cost?

Take an uncached text workload sent directly to Mistral: 10 million fresh input tokens and 2 million output tokens. Those volumes are our assumption, with no batch discount, regional surcharge, taxes or other fees; this is an illustration, not usage measured in the coding benchmark.

At Mistral's listed rates, the arithmetic is:

Input:   10M tokens x $0.70/M = $7.00
Output:   2M tokens x $2.10/M = $4.20
Total:                           $11.20

The same token mix at the Requesty catalog's listed upstream rates costs $10.98 before markup: 10M input tokens at $0.68/M cost $6.80, and 2M output tokens at $2.09/M cost $4.18. Adding the stated 5% pay-as-you-go markup gives $11.53, rounded to the nearest cent. This is a catalog-based estimate, not a verified bill; confirm the rate because Requesty's catalog and Mistral's current prices differ.

The weakest assumption is that every input token is fresh. If 5 million of the 10 million input tokens are cache reads at Mistral's listed $0.07/M, with output unchanged at 2 million tokens, the direct cost falls to $8.05: $3.50 for fresh input, $0.35 for cached input and $4.20 for output. Your cache-read share will depend on the workload.

For a real token mix, Modal reports roughly 100,000 input tokens and 500 output tokens per request for a coding-agent workload it served in production. Assuming all input is fresh and excluding fees or discounts, Mistral's Large 4 rates put that mix at approximately $0.07105 per request: 0.1M input tokens at $0.70/M plus 0.0005M output tokens at $2.10/M. Modal gives no cache-hit breakdown for that figure, and this calculation is not Modal's Large 4 bill.

Output volume and the number of attempts needed to finish a task both affect the bill. Our cost-per-task analysis explains why a lower token rate can still leave you with a higher cost per completed job.

How can you call or deploy Large 4?

Mistral lists its API and Mistral Studio as access options, alongside OpenRouter. If you have downloaded the weights, start with the Hugging Face overview and review Mistral's deployment guidance and license.

Through Requesty, Large 4 uses the OpenAI-compatible API and the managed mistral-large-4 ID. The model page also lists mistral/mistral-large-4 if you want to pin the Mistral provider.

1. 1
   Create an API key

   For Mistral direct, set your key as the MISTRAL_API_KEY environment variable. For Requesty, create a key using the quickstart, then set it as REQUESTY_API_KEY.

2. 2
   Call Large 4

   Choose a route below. The direct example uses the model ID in Mistral's chat example and its chat completions endpoint; the Requesty examples use the managed ID.

   Mistral direct: curlRequesty: PythonRequesty: TypeScript

   Shell

   Copy

   curl https://api.mistral.ai/v1/chat/completions \
     -H "Authorization: Bearer $MISTRAL_API_KEY" \
     -H "Content-Type: application/json" \
     -d '{
       "model": "mistral-large-4-0",
       "messages": [
         {
           "role": "user",
           "content": "Write a Python function that reverses a string."
         }
       ]
     }'

3. 3
   Choose a region or fallback

   If you need an EU route, review the EU routing docs and check the model's EU-only ID. You can also configure fallback policies for alternative models or credentials. The current Large 4 catalog lists one provider, so do not assume same-model failover across providers.

Which route should you choose?

Before moving a coding agent to Large 4, replay your own tasks against it and the alternatives you use today. Track completed tasks and total spend alongside token prices: the 3-point difference on our benchmark is a starting point, and your task mix and agent setup may produce a different result.