Mistral Large 4 goes open-weight — should Indian teams self-host a 1-trillion-parameter model or just call the API?

Key takeaways

• Mistral published the weights of Mistral Large 4 on 29 October 2026 under a modified MIT licence. You can run it commercially in-house; selling it as a hosted model service needs a separate licence from Mistral (model card).
• It is a ~1-trillion-parameter Mixture-of-Experts model with 49B active parameters per token, a 1M-token context window, and text + image input (Hugging Face blog).
• "Free weights" does not mean cheap to run: the reference setups use a full node of 8× NVIDIA H200 (FP8) or Blackwell GPUs (NVFP4). Even a 2-bit quantised build needs a 512 GB-RAM machine.
• Its standout areas are coding agents and defensive cybersecurity — Mistral reports 61.7% on DeepSWE v1.1 and 82% on a vulnerability reproduce-and-patch test.
• For most Indian startups and SMEs, the API is the right answer. Self-hosting makes sense only when data must not leave your infrastructure and you have steady, high token volume.

1. What actually shipped

Large 4 is served through Mistral's API as mistral-large-4. What changed last week is that the full weights are now downloadable from Hugging Face in two official formats: FP8 and NVFP4 (a 4-bit format for NVIDIA's Blackwell GPUs) (Hugging Face blog). Community GGUF builds for llama.cpp, including Unsloth's, appeared the same day.

Under the hood it follows the design most frontier open models now use:

• Mixture-of-Experts: 128 routed experts, 4 picked per token plus 1 shared expert. Only about 49B of the ~1T parameters do work on any given token, which is why it is far cheaper per token than a dense 1T model.
• Sparse attention: Multi-head Latent Attention with a DeepSeek-style sparse indexer that looks at up to 2,048 key positions per query. This is what makes a 1M-token window practical.
• Vision: a 1.6B-parameter Pixtral image encoder, so it can read screenshots, scanned forms and diagrams.
• Reasoning is always on, with low, medium and high effort settings — you cannot switch it off, so budget for the extra output tokens.

2. How good is it?

Mistral and Hugging Face quote these numbers. They are vendor-reported and have not been independently reproduced in our testing.

Read this honestly: Large 4 is not the smartest open model overall, but it is competitive on coding agents, unusually token-efficient, and among the best open options for security work. Mistral also says it refuses harmful cyber prompts more often than other open models — useful if you are putting it in front of a SOC team, annoying if you do red-team work.

The 1M-token context is the other headline. For Indian use cases that means feeding a full codebase, a year of server logs or a 600-page tender document in one go, though quality at the far end of very long contexts should always be tested on your own data.

3. What it takes to run it yourself

This is where the "free" part ends.

Sources: model card, Hugging Face blog.

The supported serving stacks are vLLM (Mistral publishes a ready Docker image), SGLang, Transformers and llama.cpp. Fine-tuning recipes exist for TRL, Axolotl and Unsloth, but the TRL example uses 8× B300 — this is not something you fine-tune on a rented single GPU.

4. The India angle: cost, data and compliance

API cost. Model trackers list Large 4's API price at about $0.75 per million input tokens and $2.30 per million output tokens (Opper tracker, reported — confirm on Mistral's pricing page before budgeting). A support bot or internal coding assistant doing 50M input and 10M output tokens a month would cost about $60, roughly ₹5,300 at ₹88 to the dollar. Because reasoning cannot be disabled, measure real output-token usage in a pilot before trusting that estimate.

Self-hosting cost. An 8× H200 node is a large fixed cost whether rented from a hyperscaler, an Indian GPU cloud, or a provider on the IndiaAI Mission compute portal. At the API price above, you would need to be processing billions of tokens a month — continuously — before a dedicated node beats pay-per-token. Few Indian SMEs are anywhere near that.

When self-hosting is still worth it:

• Data that cannot leave your premises — patient records, bank transaction data, defence or government work, or contracts that forbid third-party processors. The DPDP Act 2023 does not ban foreign processing outright, but sector regulators and client contracts often do.
• Air-gapped or sovereign environments where no external API is allowed.
• Security teams who want to run vulnerability triage on proprietary code without sending it to any vendor.
• Service companies building a dedicated on-prem AI stack for a large client — note that offering the model as a paid hosted service to many customers needs Mistral's commercial licence.

5. Large 4 versus the alternatives

Our general advice from how to choose an LLM for your product still holds: start with the API, build an evaluation set from your own tickets, code or documents, and only move to self-hosting when the numbers or the compliance requirement force it.

6. What to do this week

1. Developers: try mistral-large-4 via the API on one real task — a code-review bot or log summariser — and record accuracy and output tokens.
2. Security teams: test it on past incidents or known CVEs in your own stack; the vulnerability-patching results are its most interesting claim.
3. CTOs at regulated firms: price an 8-GPU node against your expected monthly tokens. If the break-even is years away, use a private API deployment or a smaller open model instead.
4. Students and freshers: learn vLLM, MoE basics and quantisation (FP8, NVFP4, GGUF). Serving large open models is fast becoming a distinct infrastructure skill.
5. Read the licence before building a product on it — especially if you plan to resell access.

Planning an on-prem or private-cloud LLM deployment? Our AI & automation development team can size the hardware, set up vLLM, and build the evaluation set — and our data centre & cloud team handles the GPU infrastructure underneath.