What Meta published
On October 15, 2026, Meta released Muse Spark Open, a natively multimodal mixture-of-experts model with 816 billion total parameters, 42 billion active per token, and a 1 million token context window, aimed at coding and agentic work. The Hugging Face model card describes 256 experts with 8 active per token, a 1.8 billion parameter perception encoder, and text, image, video and audio inputs with text output. Its default reasoning effort is xhigh. [1][3][2]
The weights were downloadable from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Meta also listed hosted APIs at Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI, and support in SGLang, vLLM, TensorRT-LLM, Transformers, Apple MLX, llama.cpp and Unsloth. [1]
Open, if you have a terabyte
The model card gives the memory bill: at least 700GB of GPU memory for the 4-bit NVFP4 weights and about 2TB at BF16. Meta said the model runs on two Mac Studios with Apple's M5 Ultra using the MLX framework, each with up to 512GB of unified memory, so that two machines hold the entire model. [3][1]
Atlas interpretation: Downloadable is not the same as local for most users. At this size the practical routes are a multi-GPU server, a pair of top-memory Macs, a quantized variant with its own quality trade-offs, or one of the hosted APIs. Its smaller sibling Muse Glimmer, released in August at 30 billion parameters, was the Meta model built to run on a Mac or a PC with one consumer GPU. [3][1][4]
The license is Meta's own
Meta released the weights under the Muse Model License, which it described as having limited commercial terms that let many developers use the model for free. The model card's license text sets a revenue test for anyone who sells access to AI models through an API or similar interface: if the licensee and its affiliates together take in more than $20 million in gross revenue in any 12 consecutive months, the licensee must notify Meta and sign a separate written agreement before offering the model as part of that model-as-a-service business. Internal use, research and features inside the licensee's own products are outside that test. The license also requires products to display "Built with Muse Spark Open". Downloading on Hugging Face is gated behind a form asking for legal name, date of birth and organization. [1][3]
Atlas interpretation: That is a different deal from Glimmer, which Meta released under Apache 2.0. Muse Spark Open is open weights under a custom license with commercial conditions, so its open-source tag here means downloadable weights, not a claim that the license meets an open source definition. [3][4]
Where Meta's own table puts it
Meta's announcement called it the most capable US open-weight model to date, and The New Stack framed the release as Meta's answer to China's best open-weight models. The model card's comparison table includes Kimi K3, GLM 5.3 and several US open models. Against Kimi K3 it reports Muse Spark Open ahead on JobBench (61.8 to 54.3) and MCP-Atlas (89.2 to 82.3), and behind on the Artificial Analysis Intelligence Index (39 to 44), Terminal-Bench 2.1 (76.4 to 85.0) and GPQA Diamond (89.0 to 93.5). GLM 5.3 scores 45 on the same index. Among US models in the table, Reflection AI's Beam has no index score but is ahead of Muse Spark Open on Terminal-Bench 2.1 (80.1) and GPQA Diamond (90.5). [1][3][2]
Atlas interpretation: Read together, Meta's "US" qualifier carries weight. On Meta's own card the model leads on agentic tool-use tests and trails both Kimi K3 and GLM 5.3 on the general index and Kimi K3 on coding, and a US rival beats it on some rows, so the claim rests on Meta's choice of which results count, and it is about open models from US labs, not open models overall. [3][1]
Safety claims and Meta's open-weight history
Meta said it added extra safety post-training to account for the risks of releasing weights, evaluated the model under its Meta Superintelligence Scaling Framework, and found it met the framework's release thresholds. A full evaluation report was promised for later and was not part of the release. [1]
Atlas interpretation: The release came eighteen months after Llama 4, Meta's last large open-weight family. In between, Meta launched Muse Spark as a closed model and sold access through its Model API, and opened only the small Glimmer. Meta presents Muse Spark Open as the open-weight version of Muse Spark with extra safety post-training, but did not say which closed Spark release it corresponds to. [1][3]
Sources
- Introducing Muse Spark Open, The Most Capable US Open‑Weight Model to Date
Meta AI Research · Oct 15, 2026
- Meta ships Muse Spark Open, its answer to China’s best open-weight models
The New Stack · Oct 15, 2026
- Muse-Spark-Open-816B-A42B
Hugging Face · Oct 15, 2026
- Introducing Muse Glimmer: An Open Agentic Model That Runs on Your Device
Meta AI Research · Aug 10, 2026