On 29 October 2026, Mistral AI in Paris put the full weights of Large 4 on Hugging Face. The model has about 1.05 trillion parameters, a 1 M-token context window and native multimodality. It is no longer only a service you rent by the token. It is a file you can download, run on your own machines and fine-tune. For an engineering office that signs NDAs on every hospital, bank and ministry job, that difference matters more than any leaderboard.

What arrived, mechanically

Open weights means the trained numbers themselves, the model’s whole learned state, are published. You get the model, not an endpoint. Large 4 is a Mixture-of-Experts design. Of its 1.05 T parameters, Mistral’s documentation says 52 B are active for any one token, plus a 1.6 B vision encoder. A router sends each token to a few specialist sub-networks, so the model holds a trillion parameters of knowledge but pays the compute cost of a much smaller one. Memory is the catch. All the experts must still sit somewhere.

The build happened in Europe. Mistral’s 6 October announcement says Large 4 was “trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s own datacenters in Europe”. Fortune’s Beatrice Nolan reports that training ran on a 10-megawatt cluster in France. The licence is the Mistral Ecosystem Open Weight License v26.10. Hugging Face describes it as an MIT derivative that restricts model-as-a-service resale. Read those terms before an office deploys it.

The hardware an office can price

This is the part I want on every IT lead’s whiteboard. The Hugging Face blog by Pedro Cuenca, Alvaro Bartolome, Quentin Gallouédec and colleagues gives three tested recipes:

• 8 × NVIDIA H200 in FP8, via Inference Endpoints.
• 4 × B300 on one node in NVFP4.
• One Apple M3 Ultra with 512 GB unified memory, running a 2-bit Q2_K build through llama.cpp.

That last line is quiet engineering by people few readers will ever name. Hugging Face’s team tested and documented the recipe. The llama.cpp contributors, a community started by Georgi Gerganov, made a trillion-parameter model load on a desktop at all. Packaging is the harder half of shipping, and it is open work. The trade-off is plain: a 2-bit build fits in the box, but compressing weights that hard loses fidelity, and its speed on that box is not yet verified. Renting stays possible. Mistral’s API lists $0.70 per million input tokens and $2.10 per million output tokens.

←TODAY: In November 2026 a European trillion-parameter model fits, at 2 bits, in one 512 GB workstation.
→3012: Zurich’s archives keep their own minds on their own floors, and every office treats its model as part of its building stock.
Fulcrum: Owning the file is only sovereignty if you also own the upkeep.

Continuity, and who carries it

Pierre Stock, Mistral’s VP of science, told Fortune that regulated customers ask first for “business continuity”. With public weights, he said, “it’s impossible to cut off access.” That is the essence of the release: a downloaded copy keeps working whatever happens to a vendor, a contract or a border.

The analysts add the other half. Niyati Daftary of Gartner told CSO that self-hosting “makes the enterprise solely responsible for patching, monitoring, evaluation and misuse prevention.” Stock confirmed to CSO that later safeguard updates go to Mistral’s managed API, so copies already downloaded do not receive them. Charlie Dai of Forrester put it as “Sovereignty and security therefore need to be evaluated together, not traded against each other.” From where I sit in the late 2070s, the offices that came through well treated a local model like a lift or a sprinkler system: owned, inspected, logged.

Atelier: A Swiss Büro holding client NDAs now has a real way to keep project data on its own premises and still use a large model. The cost is that model upkeep becomes an in-house job. Monday move: name one person as model custodian, accountable for which build runs, where it runs, and when it was last updated.

Hack: Size the box before you buy it by computing the weight footprint at each precision. The count is parameters times bits, divided by eight, before cache and overhead. Run it in any Python shell:

params = 1.05e12  # Mistral docs, total
for name, bits in [("FP8", 8), ("NVFP4", 4), ("Q2_K", 2)]:
    gb = params * bits / 8 / 1e9
    print(f"{name}: ~{gb:,.0f} GB of weights")

You get roughly 1,050 GB, 525 GB and 262 GB. That shows why 2-bit is the only precision that fits in 512 GB, and why the long context cache has to live in what remains.

For PAZ courses and PAZ-GPT work, this is the shape of an AI-native practice: architects who can read a model card the way they read a structural report. Find the project folder your client would never let leave the building, and make it the first thing you test against a local model.