Mistral Large 4 is free for enterprises to run inside the company and to embed inside fixed product features, and it stops being free at the point where model access is offered to third parties: the Mistral Ecosystem Open-Weight License requires any Model-as-a-Service operator to sign a separate commercial license with Mistral before commercial use [1]. Self-hosting is a separate question with its own hardware minimum. The weights went public on October 29 under that license, a modified version of the MIT license [1], 23 days after the model itself launched on October 6 [2] [3], with launch-day coverage already promising an open-weights release [4]. The release thread was posted on Hacker News on October 29 and was still on the front page a day later, with 1,654 points and 427 comments [5].
The license decides where a deployment needs a signature from Mistral, and the model card's serving examples decide what hardware to expect. Until October 29, Large 4 was served on Mistral's API [3]; the weight release is what makes self-hosting an option at all.
Internal Use and Embedded Features Carry No License Cost
The base grant is broad. Anyone obtaining the model may "use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies" and "run, deploy, further train or otherwise modify the Model and create derivative works from it" [1], and Mistral's card describes the license as covering commercial and non-commercial use alike [3]. The license's only commercial restriction is Section 2, and the text exempts internal use and any use through Mistral's official products from it [1].
Internal use is defined as "any use that does not make the Model or its derivative works available to third parties" [1]. Running the weights on company infrastructure for internal workflows, including further training on proprietary data, stays inside that definition as long as the model or its derivatives are not made available outside the company.
The Model-as-a-Service definition also excludes two product shapes on its face: "end-user products with model capabilities solely embedded within specific features or harnesses", and the "mere relaying of requests to models hosted by others" [1]. A customer-facing product can therefore stay outside Section 2's restriction when model capability is embedded inside fixed features rather than handed to customers as raw model access. The remaining question is what happens when a company starts selling that access itself.
Selling Model Access Is Where the Modified MIT Stops
Section 2 is the clause that turns the license from permissive to negotiated: "If the Licensee or any of its affiliates operates a Model as a Service business, the Licensee must enter into a separate commercial license with Mistral before using the Model or its derivative works for any commercial purpose" [1]. The definition behind it covers giving a third party access to the model or its derivatives for "inference or post-training by any means (for instance via API or hosted endpoints)", in a manner that lets that third party "exercise meaningful control over the inputs, parameters, or training data" [1].
Mapped onto product decisions, the text names its own cases. An API business or hosted endpoints where customers set prompts, parameters, or training pipelines meets the definition, and so does giving customers access to the weights to post-train the model themselves; in each case the customer exercises meaningful control over the inputs, parameters, or training data [1]. A service where the vendor post-trains the model on the customer's behalf, without the customer ever touching the model, is a different shape: the license text never uses the word "fine-tuning" and does not enumerate that arrangement. The clause binds "the Licensee or any of its affiliates", while the internal-use exemption counts third parties without saying whether affiliates are among them. The license routes these questions to Mistral's own contact form, where a commercial license can be requested [1]; the text does not settle them in advance.
The Vendor's Own Examples Run It on Multi-GPU Nodes
Staying inside those exemptions moves the constraint from licensing to hardware. Mistral's card describes Large 4 as a 1-trillion-parameter Mixture-of-Experts model with 49 billion active parameters per token, 52 billion counting embedding and output layers, built from 128 routed experts (4 active per token) plus 1 shared expert [3]. Serving examples for models of this size split one model instance across GPUs with tensor parallelism, and Mistral's examples follow the same pattern: the vLLM example launches the NVFP4 checkpoint with --tensor-parallel-size 8 [6], and the SGLang example uses --tp-size 4, which the card calls an "example tensor-parallel size for a Blackwell node" to be adjusted "to your topology" [6].
The quantized NVFP4 checkpoint pairs NVFP4 MoE experts with FP8 linear layers and targets Blackwell data center GPUs (B200, B300, GB200, GB300), which carry dedicated hardware for the format; Mistral says users should observe performance similar to the FP8 checkpoint with a substantially smaller memory footprint and faster inference [6]. GGUF conversions for llama.cpp are published under the ggml-org organization [6], and the vLLM project's recipe lists FP8, NVFP4, and an AMD Quark MXFP4 checkpoint [7].
The vLLM recipe's reference numbers indicate the class of hardware involved: on 8x GB300 with NVFP4 and the built-in DSpark draft head enabled, tensor-parallel 8 serving measured 3.1 ms per output token at concurrency 1, and a 4+4 prefill-decode split reached 241 output tokens per second per GPU at 32k input [7]. The same recipe records limits: several serving features, among them the DSpark draft head and the NVFP4 checkpoints, are not yet validated on AMD GPUs [7]. The card's deployment note adds a fallback for teams that stay local: "For optimal performance, we recommend using the Mistral AI API if local serving is subpar" [6].
One Listed API Provider, and a Contested Open Label
Artificial Analysis prices the model at $0.70 per 1M input tokens and $2.10 per 1M output tokens with a 90% cache discount [8], and lists one API provider serving it: Mistral itself [9]. Its blended price is $0.40 per 1M tokens at the panel's 7:2:1 cache-input-output weighting [9]. The panel measures the model at 161.3 output tokens per second, ranking it 11th of 78 models on that metric, and scores it 26 on its Intelligence Index, above the median of 23 among open-weights models in its size class [8]. The same panel's charts carry a separate "Open Weights (Commercial Use Restricted)" category alongside plain open weights [8], which is where the license debate lands in third-party taxonomies.
One comment on the Hacker News release thread argues, "The limits in the license mean it isn't open source" [5]; the commercial trigger sits in the license text either way, and that text is what a deploying company accepts.
The Hacker News discussion added a compliance angle. One commenter argued that because the US Department of Homeland Security already forced DeepSeek's removal from federal systems, financial companies that mirror federal compliance practice weigh a vendor's jurisdiction, and that labs headquartered in France and the UK were unlikely to face the same treatment [5]. That is one commenter's argument on the thread; it shows what procurement discussions around this release are already weighing.
The Decision Order for SaaS Selection
The license settles the first question and the model card settles the second. If the use is internal, or model capability sits embedded inside fixed product features, the weights carry no license condition beyond keeping the copyright and permission notice with copies [1]; the decision that remains is infrastructure, and the vendor's own examples put it at a multi-GPU Blackwell-class node running the NVFP4 checkpoint under tensor parallelism [6]. If the product's value is model access itself, whether as an API, hosted endpoints, or customer-driven post-training, Section 2 applies and the commercial license conversation with Mistral comes before launch [1].
The buy option, meanwhile, is a single vendor: Mistral's own API is the only provider Artificial Analysis lists [9]. For internal and embedded uses the license fee is zero, so the decision that gates self-hosting is hardware capacity; for a business whose product is model access, it is the commercial license [1]. The release made the weights available to any company's infrastructure; the license decided which uses of them are free and which need Mistral's signature.