The Pulse
Meta Releases 816B-Parameter Muse Spark Open Weights
Meta released Muse Spark Open, an 816-billion-parameter model with 42 billion active parameters and a one-million-token context window. The announcement includes a Key Benchmarks section and links to an evaluation methodology report.
AI.info Team · Written with AI ·
Meta’s new Muse Spark Open has 816 billion parameters, but only 42 billion are active for each token. The company released the model’s weights on October 15, alongside a one-million-token context window and support for text, image, audio and video inputs. Meta says the model targets coding and agentic work, including tasks that require extended reasoning and multiple steps.
The release makes Muse Spark Open available through hosted inference providers and for download from Hugging Face and cloud model catalogs. Developers can run it with inference software including SGLang, vLLM, TensorRT-LLM, Transformers, MLX and llama.cpp. Meta describes it as the most capable U.S. open-weight model to date; its announcement includes a “Key Benchmarks” section and links to an evaluation methodology report.
Open weights, with terms attached
“Today, we’re releasing the weights for Muse Spark Open, the most capable US open-weight model to date.”
Meta AI Research writes. Meta AI Research
Meta distributes the model under its Muse Model License, which it describes as having limited commercial terms that allow many developers to use it for free. That is an open-weight release, not an unrestricted release of every part of the model’s development. Meta also says the model received additional safety post-training and met the release thresholds in its Superintelligence Scaling Framework; it points readers to a model card for details and says a full evaluation report is forthcoming.
The distinction matters for teams deciding whether they can adapt, host and deploy the model on their own terms. Meta’s developer documentation says use is also subject to an Acceptable Use Policy. The announcement does not spell out those restrictions in detail, so developers will need to review the license and policy before relying on the weights in a commercial product.
816 billion parameters, and a sizeable hardware bill
The model uses a sparse mixture-of-experts design: each token is routed to a subset of the model, so computation tracks the 42 billion active parameters. But serving still requires loading the full 816-billion-parameter set into memory. Meta’s documentation lists a BF16 checkpoint at about 1.65 terabytes and a compressed NVFP4 version at about 537 gigabytes; the BF16 version can require roughly two terabytes of GPU memory once cache, activations and runtime overhead are included.
That footprint puts local deployment far beyond an ordinary desktop. Meta says the model can run on two Mac Studio systems with M5 Ultra chips using Apple’s MLX framework, each with up to 512 gigabytes of unified memory. For GPU deployments, the company’s published configurations include eight B300 GPUs for BF16 and a four-GPU Blackwell system for NVFP4. Hosted providers offer another route for developers who do not have that hardware.
Meta’s desktop demos focus on code and video
Meta says it tested on-device workflows through Muse Code, its coding tool powered by Muse Spark Open. In one example, the model made bug fixes and feature changes to Ice Cubes, an open-source Mastodon client, then completed a build. In another, it operated DaVinci Resolve and reviewed footage frame by frame to identify moments and crop them vertically.
The company also reports custom, auto-tunable Helion GPU kernels for SGLang. Meta says those kernels delivered 1.03 to 3.95 times faster performance at the kernel level on NVIDIA and AMD GPUs, while end-to-end speed increased by 3% on AMD MI350. Kernel-level gains measure a component of inference, not the full application experience; the smaller end-to-end figure gives a more direct indication of the benefit in that specific test.
Availability spans hosted APIs and downloadable checkpoints
At launch, Meta lists Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI as inference options. It also lists Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face and Microsoft Foundry as places to access the weights. The documentation identifies the model as multimodal, with text output and text, image, audio and video inputs, and says training data covered more than 100 languages.
Meta’s release pairs broad access to the weights with substantial hardware requirements and a license that carries conditions. For individual developers, the announced two-Mac-Studio setup is the clearest local path; for many others, the practical options are a compressed checkpoint on a multi-GPU server or a hosted API.