Back to Blog
    AI & AutomationNov 17, 2026

    Mistral Large 4 Is Now Generally Available: Self-Hosted Frontier AI for Business

    Amir

    Amir Farid

    7 min read

    Mistral Large 4 Is Now Generally Available: Self-Hosted Frontier AI for Business

    Mistral Large 4 Is Now Generally Available: Self-Hosted Frontier AI for Business

    Mistral Large 4 is now generally available: version v26.10, dated 5 November 2026 on Mistral's own model page — a month after the preview opened on 6 October, a week after the open weights landed on Hugging Face on 29 October. It is a 1-trillion-parameter multimodal model you can download and run yourself, with a 1-million-token context window, at roughly half the preview price.

    If you read our 30 October guide to the open-weights release, this is the follow-up: the preview is now supported, two older models are retiring and the price has changed. Here is what that means if you want to self-host a frontier model.

    What "generally available" actually changes

    General availability is when a model becomes safe to build a business process on. During a preview, the API, weights and pricing can all move under your feet. From 5 November, Large 4 is version v26.10, the supported release, at the published list price.

    Mistral also says post-preview training lifted Terminal-Bench 13 points and cut output tokens 30%. Those are Mistral's own claims, not independently verified — but fewer output tokens means lower bills. Treat them as a reason to test: our GDPx coverage notes no model cleared roughly 12% on long-horizon tasks.

    The model in plain English

    Server racks with orange cabling and a glowing network diagram in a data centre

    Large 4 is a sparse mixture-of-experts model: about 1 trillion parameters in total (1.05T), with only around 52 billion active per token. A dense model would use every parameter for every token; a mixture-of-experts routes each token to a few specialist sub-networks, so you get frontier-scale capability far cheaper to run. Mistral calls it its largest model to date, trained in its own European data centres.

    It takes text and image input, produces text output, and offers a 1-million-token context window — a contract pack or a large codebase in one call. It is, Mistral says, fluent in 160+ languages. Reasoning effort is configurable — low, medium or high — and function calling plus strict JSON output suit AI agents.

    Where Mistral says it is strongest

    Mistral positions Large 4 as a generalist strong in cybersecurity, finance and law, claiming it beats closed models on visual grounding. The model card reports 48% on the Artificial Analysis Coding Agent Index, effectively tied with the leading open-weight model, using 47% fewer tokens per task. See our Qwen4-Max coverage — a flagship you cannot download at all.

    What it costs now

    On the Mistral API, Large 4 now costs $0.70 per million input tokens, $0.07 per million cached input tokens, and $2.10 per million output tokens — down from $1.36 / $4.18 in the preview, roughly half price. Cached input matters most for agents: re-reading the same prompt and history every turn costs $0.07 per million tokens — effectively free.

    But that is only one of two prices. Self-hosting swaps per-token fees for hardware and engineering: heavy, steady workloads usually favour self-hosting; spiky ones favour the API. Judge cost per completed task, not per token — our small-business AI agents guide explains how to scope that pilot.

    API or self-hosted? A quick comparison

    • API: $0.70 / $0.07 cached / $2.10 per million tokens; no hardware; your data leaves your network.
    • Self-hosted: weights free; you pay for GPUs, hosting and maintenance; data stays inside your boundary.
    • Hybrid: prototype on the API, move steady workloads in-house once volume justifies it.

    The deprecation deadline: two models retire on 30 April 2027

    Mistral Medium 3.5 and Mistral Large 3 were deprecated on 31 October 2026; their endpoints retire on 30 April 2027. Mistral's own models page names Large 4 as the replacement for both — just under six months away.

    If either model is in production, treat this as a migration, not a configuration change. Old prompts will not transfer perfectly, and old evaluations need re-running. Start with an inventory: search your codebase and AI automation workflows for mistral-medium-3-5 and mistral-large-2512. Then run real tasks on Large 4 beside the old model, re-tune prompts, and set a switch-over date well before 30 April 2027.

    The licence: read it before you deploy

    Large 4 is released under the Mistral Ecosystem Open-Weight (MEOW) licence, a modified MIT licence. Commercial use is allowed — build products, fine-tune, run it for paying customers. But one hard limit is stated plainly on the Hugging Face model card: selling access to the model itself as a service needs a separate commercial licence from Mistral.

    Check that with your lawyer first. For deployment, Mistral recommends vLLM for production serving, with SGLang, Transformers and llama.cpp also supported; an NVFP4 checkpoint is available for smaller footprints. Our NVIDIA NemoClaw guide covers governed agents on your own servers; our self-hosted Claude managed agents coverage shows the closed-model equivalent.

    Self-hosting: what it really takes

    Downloading the weights is easy; running a 1-trillion-parameter model in production is a serious infrastructure job. The reference vLLM setup uses 8 GPUs in parallel; the NVFP4 checkpoint exists because the full model is heavy. You need multi-GPU servers and someone to keep them healthy — or a managed host.

    So the hybrid path wins for most businesses: start on the API, prove the use case, measure volume, then decide whether self-hosting pays. The 1-million-token context window can change your architecture, but one bad prompt can burn a lot of tokens, so guardrails matter more, not less. For customer-facing chat, read our AI chatbot vs AI agent guide; for internal knowledge, our RAG chatbot guide covers chatting with your own company data safely. See also our DeepSeek SuperAgent guide.

    When self-hosting is clearly the right call

    • Your data cannot leave your network — regulated finance, legal, healthcare or government work.
    • Your steady volume makes per-token fees cost more than your own GPUs.
    • You need fine-tuning on proprietary data you will not send to a vendor.

    Why Mistral is betting big — and why it matters to you

    Mistral is committing roughly €4 billion to data centres in France and Sweden and targeting €1 billion in 2026 revenue, per today's AI news briefing. Supplier longevity matters when you standardise on a model — it needs a vendor still supporting it in two years.

    There is also a sovereignty angle: a European model, trained in European data centres, with weights you can run yourself, differs from an API-only model hosted abroad — the same logic behind the du National HyperAI platform in the UAE.

    What your business should do this week

    Do three things: check for deprecated models, run five real tasks, price both routes.

    Step 1: check for deprecated models

    Search your code and contracts for mistral-medium-3-5 and mistral-large-2512. If either appears, you have a migration due 30 April 2027 — put it on the roadmap now.

    Step 2: run five real tasks on Large 4

    Pick five real jobs — a support reply, a document summary, a code review, a data extraction, a draft email — run them on Large 4, then on your current model, and record success, tokens and time. Our custom software development team builds these pilots; our London hiring guide and UK cost guide help you budget one.

    Step 3: price both routes

    Compare the API bill at $0.70 / $2.10 per million tokens against hosting it yourself, including engineering time — the cost most teams forget. If the numbers are close, stay on the API and revisit in six months; if self-hosting wins clearly, pilot the NVFP4 checkpoint. See our Gemini 4 Carbon coverage for regulated work.

    Frequently asked questions

    Is Mistral Large 4 free to use?

    Weights are free to download under the MEOW licence; commercial use is allowed. But running it costs money — API fees or your own GPUs — and selling access to it as a service needs a separate commercial licence.

    When did Mistral Large 4 become generally available?

    On 5 November 2026, as version v26.10, per Mistral's own model page. Preview opened 6 October 2026; open weights followed 29 October 2026.

    Which Mistral models are being retired, and when?

    Medium 3.5 and Large 3 were deprecated on 31 October 2026; their endpoints retire on 30 April 2027. Large 4 is the replacement for both — migrate well before the deadline.

    Can I really run a 1-trillion-parameter model myself?

    Yes, but it is a serious infrastructure job: the reference vLLM setup uses 8 GPUs in parallel. Prototype on the API first; self-host once volume justifies it.

    #Mistral Large 4#open-weight AI#self-hosted AI#AI agents