Meta has handed developers a very large model they can download, and attached a license that is not quite open source. Muse Spark Open, released on October 15, 2026, has 816 billion parameters, of which 42 billion are active for any given request, and a one-million-token context window. Three days earlier, on October 12, the company launched Muse Wonder, a more capable model that stays closed. The pairing shows how Meta now intends to compete: keep the best model proprietary, give away the next tier down.
Open weights, with conditions
Meta’s research blog says the weights ship under the Muse Model License, which it describes as having “limited commercial terms” that let many developers use the model for free. The weights are available through Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry, with hosted APIs from providers including Together AI, Fireworks, Groq and OpenRouter. Meta says it also runs on common inference tools such as vLLM, SGLang, llama.cpp and Apple’s MLX.
The license terms are where the details live, and reports are not entirely consistent. Trending Topics, summarizing the license, says users may fine-tune, distill and redistribute the model and that Meta claims no rights in outputs. It says anyone offering the model as an AI service with more than $20 million in revenue over 12 months must give Meta advance notice and sign a separate agreement, while internal use, research and embedding the model in your own products are outside that threshold. Another secondary report, however, described the threshold as $20 million per month across a company’s entire business. Those are materially different, so anyone near the line should read the license text itself rather than rely on summaries.
There is also an attribution requirement. Products using the model or a derivative must prominently show a “Built with” notice; Trending Topics gives the wording as “Built with Muse Spark Open,” though another report gave a shorter phrase. The acceptable use policy reportedly bans weapons development, malware, critical-infrastructure operation, fraud and impersonation, and the model is not intended for users under 18. Meta’s blog adds that the model got additional safety post-training because it is open-weight and that a full evaluation report is still to come.
What it takes to run
Size is the real gate. Trending Topics, citing Unite.AI, says the model needs at least 700 gigabytes of graphics memory depending on format. Meta’s own claim is that it is the first model of its class that can run on a desktop, using two linked Apple Mac Studios with M5 Ultra chips. The same report puts the cost of that setup at roughly that of a compact car. For nearly every startup, the practical route is a hosted API rather than self-hosting.
Meta’s blog gives no benchmark figures in the section that would normally hold them, so how Spark Open compares to rivals is not established by the primary source. Kernel-level speedups of 1.03 to 3.95 times on NVIDIA and AMD chips are reported for custom code in SGLang, but that is an engineering result, not a quality measure.
The strategy behind it
TechCrunch’s report on Muse Wonder frames it as Meta’s first frontier model since Alexandr Wang restructured the company’s AI efforts in August 2025. Meta says the Wonder models score comparably to OpenAI’s GPT-6 Astra and another leading rival on its own testing, and that Muse Wonder costs $4 per million input tokens and $20 per million output tokens through the Meta Model API, roughly in line with GPT-6 Sol and Opus 5.5. Wonder Full is limited to Muse Code subscribers and is not offered through the API. Meta’s emphasis is “active memory”: the ability to remember what matters to a user over days, weeks and months.
The open release is a reversal. Trending Topics notes that Meta’s spring Muse Spark launch was closed-weight, breaking from the earlier Llama approach, and that the new license resembles Moonshot AI’s terms for Kimi K3. Chinese labs such as Alibaba and Zhipu have mostly used permissive licenses without revenue thresholds or naming requirements. In effect, Meta is competing for the open-weights audience while keeping tighter control over who can resell the model, which mainly affects large cloud and AI providers rather than ordinary developers.
What This Means for Philippine Founders
For a typical Philippine startup, the license is unlikely to be the obstacle. A company building its own product on top of the model, rather than reselling it as an AI service, falls on the permissive side of the revenue clause as reported, though the monthly-versus-annual discrepancy above is worth settling by reading the actual license. The visible costs are the “Built with” notice in the product and a review of the acceptable use rules, particularly for fintech, ads or content products that touch fraud or impersonation.
The more important question is whether the model is useful enough to justify the hosting. Because self-hosting needs hundreds of gigabytes of GPU memory, most teams will call it through a provider, which makes it a pricing and quality comparison against closed APIs rather than a way to avoid vendor dependence. A short pilot on real tasks, including Filipino and Taglish text whose performance the sources do not cover, will say more than any announcement. Teams with data-residency needs should also check where each hosting provider runs inference before committing.
Share this article