Open-weight AI catches up: Muse Spark Open and Mistral Large 4

"Open weights" means the trained model itself, the large file of numbers that makes it work, is published for anyone to download and run on their own hardware. That is different from using a model through someone else's API. You control where the data goes, you can fine-tune it, and nobody can switch it off or change it under you. October 2026 brought two major releases of this kind from Western companies.

Meta Muse Spark Open

On 15 October Meta released the weights for Muse Spark Open, which it describes as "the most capable US open-weight model to date". The key specifications, according to Meta:

• a natively multimodal model with 816 billion total parameters, of which 42 billion are active for any given token;
• a 1 million-token context window;
• built for coding and agent work, especially long, multi-step tasks.

The "total versus active" split is the design trick that makes this practical. Models like this are built from many specialist sub-networks, and only a small share of them is used for each word. The model needs a lot of memory to hold, but much less computation per token.

On your desk

Meta's headline claim is that Muse Spark Open is the first model in its class to run on a desktop setup: two Apple Mac Studios with M5 Ultra chips, using Apple's MLX framework. Meta explains that each machine can give its GPU up to 512 GB of unified memory, so two of them hold the whole model, and that the chip's 1.2 TB/s of memory bandwidth is fast enough for the small slice each token needs.

Meta says it tested agent workflows running entirely on the device using its Muse Code tool. Examples include shipping bug fixes and features to an open-source iOS app, and driving the DaVinci Resolve video editor from start to finish.

If you don't want to run your own hardware, Meta lists hosted access through providers including Together AI, Fireworks, Groq, Baseten, Modal, Nebius and OpenRouter. The weights can be downloaded through Hugging Face and the major cloud marketplaces.

The licence matters

Muse Spark Open is released under the Muse Model License "with limited commercial terms", which Meta says "allow many developers to use the model for free". That is not the same as a standard open-source licence. If you are a business, read the licence before building on it, especially if you expect to grow large.

Mistral Large 4 ("Le Chonk")

On 6 October Paris-based Mistral launched a public preview of Mistral Large 4, nicknamed, in Mistral's words, "very officially: le Chonk". The specifications, according to Mistral:

• about 1 trillion parameters, 52 billion active, natively multimodal;
• trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European data centres;
• multilingual training data covering more than 160 languages, including every official EU language;
• preview access now through Mistral Studio, with weights promised "by the end of the month".

Mistral frames the release around sovereignty: a European deployment that Mistral operates end to end "under European law", and weights that organisations can run on their own servers. The work is funded by a €3 billion Series D, which Mistral calls the largest equity round ever raised by a European technology company.

The cyber argument

Mistral makes an unusually direct argument about cybersecurity. It points out that leading closed models, naming Claude Opus 5.5 and GPT-6 Astra, score near zero on one vulnerability-reproduction test "because they refuse to perform the task", and that defenders need tools that are not blocked by those refusals. Mistral says ML4 scores 82% on that test. It also says the model's refusal rate on malicious cyber prompts is higher than that of other open models.

This is where the debate gets uncomfortable. Closed labs restrict cyber capabilities because those capabilities can be misused. The OpenAI–Hugging Face incident showed what capable agents can do. Once open weights are published, a developer cannot take them back or add a restriction later. Mistral's response is to red-team the model with cybersecurity leaders, vetted partners and state authorities before the weights are released. Whether that is enough is a fair question to ask.

Safety in open models

Both companies address safety. Meta says Muse Spark Open has "additional safety post-training" because the weights are open, and that it met the release thresholds in its Meta Superintelligence Scaling Framework, with a full evaluation report to follow. Mistral reports strong resistance to prompt-injection attacks on public benchmarks. These are the companies' own assessments. Independent evaluations will matter a great deal.

What it means for UK businesses

• Data control: running a model on your own hardware keeps sensitive data inside your organisation, which can help with GDPR and client confidentiality.
• Cost: two high-end Mac Studios are a serious purchase, but they are a one-off cost rather than a monthly API bill. For heavy, steady use the numbers may work.
• Licences: check the Muse Model License terms, and wait for Mistral's final licence when the weights are released.
• Capability gap: the companies' own benchmarks put these models close to, but generally behind, the best closed models. For many everyday tasks that will not matter.
• Responsibility: when you run the model yourself, you also take on the safety and misuse controls a provider would otherwise handle.

The bottom line

The gap between closed frontier models and what you can download is narrowing. Muse Spark Open shows a very large model running on desktop hardware. Mistral Large 4 shows Europe building frontier-scale models on its own infrastructure. The trade-off is now easier to see: open weights give you control and independence, and they leave the responsibility for safe use with you. For the wider model picture, see AI Models, and for the chips behind it, Hardware & Industry.