Muse Spark Open is Meta's new open-weight AI model, released on Thursday 15 October 2026 (US time). Meta AI chief Alexandr Wang called it "the most capable US open-weight model to date", and Meta says it is built for coding and agentic work. The headline twist is hardware: Meta worked with Apple so the full model can run on two Mac Studios with M5 Ultra chips. Here is what the model is, how big it is, how the licence works and how it compares with the Chinese open models it is chasing.
What is Muse Spark Open?
According to AFP's report in The Guardian, Muse Spark Open is a downloadable model whose weights developers can inspect and modify, unlike closed systems such as ChatGPT or Claude. Meta says in its blog post that the model "excels at long-horizon tasks, complex software engineering, and multi-step reasoning" and ships with "strong safeguards". It arrives days after Meta announced a new closed frontier model, so Meta now runs two tracks: a closed flagship and an open model for developers.
Muse Spark Open specs
AI News Online's 16 October briefing, which cites Meta's model card, lists:
- Architecture: mixture-of-experts (only part of the network is active for each token)
- Size: 816 billion total parameters, 42 billion active
- Context window: 1 million tokens
- Modality: described by Meta as natively multimodal
- Availability: weights on Hugging Face, with hosted access through providers including Together, Fireworks, Groq and Baseten
The Apple tie-up: running it on two Mac Studios
The Guardian reports Meta says the model is "optimized" for Apple's M5 Ultra chips and can run on two linked Mac Studios, effectively a desktop "supercomputer". AI News Online adds that the setup uses Apple's MLX framework and two 512GB M5 Ultra Mac Studios, and that MacRumors reported the 512GB configuration is not yet on sale and is expected later this month. That briefing also notes the 256GB Mac Studio alone costs over $10,000, so a two-machine rig is a serious studio or lab purchase, not a hobby box.
Muse Spark Open licence: what you can and cannot do
The model ships under a Muse Model License, not a standard open-source licence. Per AI News Online, citing Trending Topics, it requires a "Built with Muse Spark Open" attribution, and companies with more than $20 million in annual revenue that serve the model as an AI service need a separate agreement with Meta. For Indian start-ups and solo developers below that line, that means free download and self-hosting with attribution; larger platforms should read the terms before shipping it to customers.
People also ask: is Muse Spark Open better than Kimi K3 or GLM?
Not across the board, on Meta's own numbers. The Guardian notes Chinese models such as Alibaba's Qwen, Moonshot AI's Kimi K3 and Z.ai's GLM have led open-weight AI for the past couple of years, and Meta says it now competes at a similar level. AI News Online's reading of the model card:
- Ahead of open rivals on agent benchmarks such as MCP-Atlas (89.2) and JobBench (61.8)
- Behind Kimi K3 and GLM 5.3 on coding tests like Terminal-Bench 2.1 (76.4 vs 85.0)
- Behind on the Artificial Analysis index (39 vs 44 and 45)
All of these figures are vendor-reported, so wait for independent tests before treating any ranking as settled.
Why this matters
Three things stand out. First, a US lab is back in the open-weight race at frontier scale after a period when Chinese labs set the pace. Second, the Apple partnership makes a very large model runnable in an office without a data centre, as long as you can afford the hardware. Third, the licence shows the new normal for "open" AI: free weights with commercial strings attached for the biggest users.
For background on the closed-model race, read our GPT-6 Astra explainer, and for another view of China's AI push see our Tencent Gander AI explainer.
Sources
The Guardian / Agence France-Presse, "Meta releases Muse Spark Open, a new 'open-weight' AI model" (15–16 Oct 2026); AI News Online daily briefing, 16 Oct 2026 (citing Meta's model card, Hugging Face, CNBC, MacRumors and Trending Topics).