In the space of two weeks this October, Meta, DeepSeek and Mistral each published AI models anyone can download. One of them, Mistral Large 4, has about a trillion parameters and is what Mistral calls its flagship. You could download any of them today. Almost no small business should. It is still worth knowing why, and what these releases mean for the AI prices you pay.
What "Open Weights" Means
An AI model is, underneath, a very large file of numbers called weights, which the model learned during training. Parameters are those individual numbers; more usually means more capable and costlier to run. A company that releases open weights publishes that file, so anyone can run the model on their own computers or pay any hosting company to run it for them.
A closed model works the other way: you only rent it, through the maker's app or its API (the connection other software uses to send it work, usually billed per token, a small chunk of text). Meta now does both. Its new Muse Wonder is closed; VentureBeat reported that "Neither Wonder model is being released with publicly downloadable weights," at $4 per million input tokens and $20 per million output tokens.
Open weights is not quite open source. You get the finished model, not necessarily the training data, and a license sets the rules. Licenses vary a lot.
October's Three Releases
- Muse Spark Open (Meta, October 15). 816 billion parameters, of which roughly 42 billion work on each token. That design, called mixture of experts, sends each piece of text to a few specialist parts of the model, which makes it cheaper to run. It accepts text, image, video and audio, with a 1-million-token context window (how much it can read at once). It ships under Meta's own Muse Model License, and Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI host it.
- DeepSeek V4.1-Flash-1017 (DeepSeek, mid-October). A "minor update" to its Flash model, with weights under the MIT license, one of the most permissive there is. SiliconANGLE reports API prices are unchanged; at the September launch, off-peak rates started at 15 cents per million uncached input tokens. In DeepSeek's own testing it solved 77% of tasks on a coding test called DeepSWE v1.1, up from 74.2%.
- Mistral Large 4 (Mistral, October 29). Unveiled October 6, when TechCrunch reported it was reachable only through a public guardrail endpoint (an access point with safety filters), and opened up about three weeks later. SiliconANGLE reports 1 trillion parameters with 52 billion active. Its license is "a modified MIT license" with one condition: businesses that operate a Model-as-a-Service offering, meaning they sell access to the model itself, need a separate commercial license. Mistral's own API charges $0.70 per million input tokens and $2.10 per million output tokens.
Why You Probably Should Not Run One Yourself
Meta's own guide is a reality check. Even compressed, Muse Spark Open is about 530 GB and serves on four Blackwell Ultra GPUs or eight H200s, both top-end data center chips. The desk-sized option is two Mac Studios with M5 Ultra chips and 512GB of memory each. Then someone has to keep it updated and secure. For a five-person firm drafting proposals, that is the wrong tool.
When Open Weights Do Matter
Most small businesses end at the last box.
Even if you never download a model, open weights help you in two ways:
- Price pressure. Mistral's $0.70 and $2.10 per million tokens are a fraction of Muse Wonder's $4 and $20. These are list prices, not a quality comparison. We covered the wider trend in AI got cheaper.
- Choice of host. Because anyone can run the weights, several companies compete to sell the same model, at least seven for Muse Spark Open alone. Choose one by price, location or data terms, and switch if it changes.
Two cautions. Read the license before building a product on any of these, especially if you resell AI features. And check where a host processes your data before sending it anything sensitive; DeepSeek, for example, is a Chinese company based in Hangzhou, and open weights mean you can choose a host whose data terms you have checked. Our guide to where your data goes in AI tools covers the questions to ask.
The Bottom Line
Open weights matter hugely to the AI industry and only indirectly to most small businesses, through lower prices and more choice. If you have a strict data rule or one high-volume job, they deserve a test. Otherwise, pick the model that fits how you already work and move on.
Sources
- Meta, "Build with Muse Spark Open" (October 15, 2026): model details, license, hosts, hardware. dev.meta.ai
- SiliconANGLE, "DeepSeek releases updated V4.1-Flash model it says completes tasks faster" (October 18, 2026): license, pricing, test results. siliconangle.com
- TechCrunch, "Mistral's new 1T model aims to leapfrog closed and open rivals" (October 6, 2026): preview-only access. techcrunch.com
- SiliconANGLE, "Mistral releases open weights for Mistral Large 4, providing sovereign control for the company's 'Le Chonk' AI model" (October 29, 2026): weights release, 1 trillion total and 52 billion active parameters. siliconangle.com
- Mistral AI, "Mistral Large 4 (Le Chonk) is now open weight" (archived October 29, 2026): flagship model, license terms, API pricing. web.archive.org
- VentureBeat, "Meta's new Muse Wonder and Muse Wonder Full models achieve frontier performance, but only the weaker one is available via API" (October 12, 2026): Wonder's closed weights and API prices. venturebeat.com
Have a data rule or a high-volume AI job?
We will test an open-weight model against what you use now on your own examples, and tell you plainly whether it is worth switching.