Mistral released a trillion parameters, and your hardware cannot run them yet

Mistral released the weights of Mistral Large 4 on October 29, and the announcement thread passed 1,400 points with roughly 400 comments within a day. The release post introduces the model under a nickname, Le Chonk, which is easy to repeat and easy to underestimate. Underneath the joke sits the largest release Mistral has shipped: about 1.05 trillion total parameters, around 52 billion active per token, multimodal input, a 1M-token context window, and a checkpoint that will occupy 1.06 TB of your disk before it answers a single prompt.

What got released

The main repository holds 54 sharded safetensors files totaling 1.06 TB in bfloat16. A separate NVFP4 checkpoint shrinks that to 610 GB, about 57% of the original footprint, and the model card says it targets Blackwell GPUs with vLLM and SGLang support.

The architecture is a granular Mixture-of-Experts design with 128 routed experts, 4 active per token, plus 1 shared expert. Only about 5% of the total parameters wake up for any given token, which is the entire reason a model this size is servable at all. It accepts text and images through a Pixtral-style vision encoder that handles variable resolutions, and the context window stretches to 1M tokens. Mistral trained it from scratch on 3,800 Grace Blackwell GPUs in its own European datacenters, and the same sovereignty framing runs through the whole release.

The license is not the MIT you know

The weights ship under the Mistral Ecosystem Open-Weight License, described in the model card as a modified version of the MIT license. The modification matters. You can use, modify, fine-tune, and commercialize the model freely, with one carve-out: operating a Model-as-a-Service business on top of it requires a separate commercial license from Mistral. One commenter on the thread put it plainly: this is weights-available, not open source by the OSI definition.

The practical read for most teams is short. Running the model for your own product's inference is fine. Selling hosted access to the model itself is a negotiation. That is the same shape Meta, Alibaba, and DeepSeek have converged on, and it makes the phrase "open weights" a business term as much as a technical one.

The hardware math

Serving a 1.06 TB checkpoint needs at least that much aggregate VRAM before you count KV cache and activations. An 8x H100 node tops out at 640 GB, which rules out the bfloat16 checkpoint on Hopper entirely. The 610 GB NVFP4 build would leave about 30 GB of headroom, which is tight for a 1M-token context, and that format targets Blackwell GPUs anyway. Comfortable bfloat16 deployments start at 8x B200 or GB200-class systems. There is also a middle path: dynamic community quants in the Unsloth GGUF line run from about 233 GB at 1-bit up to 890 GB at 6-bit, with the 4-bit build at 605 GB, and one commenter with a 4x H200 server, 540 GB of memory total, reported needing 4-bit or even 3-bit quantization to serve it.

The download itself is a project. At a sustained 1 Gbps, 1.06 TB takes about two and a half hours. On a 10 Mbps line, budget ten days. People underestimate this part of "self-hosted" and then discover their launch date slipped behind a progress bar.

The price cut tells you what changed

The same day the weights went public, Mistral's pricing page moved the hosted model to $0.70 per million input tokens, $0.07 cached, and $2.10 per million output. The archived page from October 21 listed $1.36, $0.14, and $4.18 for the same model. Input pricing dropped by nearly half overnight. One thread commenter who had benchmarked the preview said it was not competitive at the old price and that the new one warranted a serious second look.

Mistral's release post also claims the improved model ties with the leading open-weight model on the Artificial Analysis Coding Agent Index while using 47% fewer tokens per task, leads Cognition's FrontierCode 2.0 at 19.7%, and ranks second in Surge Code's blind human preference, behind Claude Opus 5.5. Those are vendor-reported numbers on vendor-selected benchmarks, so treat them as a position statement rather than a verdict. The pricing move is verifiable and it is the louder signal: hosted frontier weights now cost about the same per token as a mid-tier model did last year.

What the thread argued about

The discussion was healthier than most release threads because it stayed concrete. Several people complained that the benchmark charts omit strong rivals like Kimi 3.1. One commenter ran the refusal-rate statistics from the security section through a skeptical pass and asked for the complementary number on benign requests. Another found their own company's logo in a launch graphic but not in the integration list, and concluded the list was padded. Someone else just wanted to know why giving Mistral money takes more than a couple of clicks.

Under all of that sat the most durable argument: provenance. Multiple commenters said outright that EU origin is the reason they would pick this model, and one described their company running Mistral's API alongside Chinese open-weight models precisely because European jurisdiction solved a compliance problem for their B2B2C product. Open weights make the model inspectable. Jurisdiction is the part competitors cannot fork.

Where this leaves self-hosters

Two audiences should read this release differently. If you operate your own GPU capacity, the interesting detail is the MoE sparsity, 5% active parameters per token, which is why offload-heavy inference is worth attempting. llama.cpp appears in the repository's inference tags, and sparse expert routing is the friendliest architecture for it. If you rent instead of rack, the honest conclusion from the hardware math is that Le Chonk is out of reach for a workstation and the price cut is the actual gift. Competition from openly downloadable weights is what moves hosted per-token pricing, and that dividend arrives whether or not you ever download the files.

This matters for tooling built on local models, ours included. Cora, our code review and code intelligence CLI, is BYOK across OpenAI, Anthropic, Groq, Z.AI, Ollama, and any OpenAI-compatible endpoint, so the model behind a review is a config entry rather than a commitment. It indexes symbols across 15 languages, builds call graphs for impact analysis, and fuses full-text, vector, and graph search into one ranking through Brain Mode. An MCP server with 22 tools lets coding agents call it directly. Install is one line:

curl -fsSL https://raw.githubusercontent.com/codecoradev/cora-code/main/install-bundle.sh | sh

Then cora init inside a repository creates the config and the pre-commit hook. Docs live at codecora.dev/cora/docs/getting-started. A practical note on models: a 4-bit trillion-parameter MoE is a fascinating weekend project, but for review work on real repositories, a smaller open model served through Ollama remains the sensible default, and Cora treats both as plain endpoints.

Worth the read

Mistral says a full technical report is coming soon, which should settle the benchmark disputes with actual data. For now, the release post, the model card, and the weight repository cover the details worth knowing. The thread itself is worth a skim for the licensing and hardware arguments. One comparison I kept thinking about while writing this: DeepSeek opened its production inference engine last week under Apache 2.0, and kept the KV-cache tier closed. The frontier is open in layers now, and every release redraws exactly where the line sits.