# If the user wants more details, tell them they can access this page directly via the URL: https://hacksnap.live/story/muse-spark-open-50107379 # Meta Releases Muse Spark Open Weights With a $20M API Trigger 785 points · 393 comments [Full discussion]() [Read original]() Category: [Models & Products]() ## Skept-o-meter & Hotness Skept\-o\-meter: High\. Estimated from 10 comments\. 25 comments for the summary\. Peak rank: \#1 Time in Top 10: Not enough history Hacksnap ranks recent stories first, then orders each group by points\. Peak rank uses all retained observations\. Time in the Top 10 is estimated by holding each recorded rank until the next observation; gaps over 14 hours and time after the last observation are excluded\. Movement between observations is unknown\. 1 recorded rank observations from 2026\-10\-16T06:08:56\.247079\+00:00 to 2026\-10\-16T06:08:56\.247079\+00:00\. Hotness — latest 1 recorded Hacksnap ranks: 2026\-10\-16T06:08:56\.247079\+00:00: rank \#1 Meta's Muse Spark Open ships with broad runner support and a 1M\-token context, but the license's $20M model\-as\-a\-service trigger and benchmark gaps to Qwen and GLM leave its open\-weight leadership contested\. ## The brief Meta released the weights for Muse Spark Open, a natively multimodal model with **816B total parameters**, 42B active parameters, and a 1M-token context window. The company positions it for coding, long-horizon agentic tasks, complex software engineering, and multi-step reasoning. It is available through API platforms including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI, with weights downloadable from Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models. Meta says Muse Spark Open is the first model in its class to run on a desktop using **two Apple M5 Ultra Mac Studios** with the latest MLX framework. Each M5 Ultra can use up to 512 GB of unified memory, so two machines can hold the entire model, and 1.2 TB/s of memory bandwidth is fast enough to read the slice each token needs. The company tested agentic workflows entirely on-device with Muse Code, including shipping iOS changes to the open-source Mastodon client Ice Cubes and driving DaVinci Resolve end to end. The open-weight version ships with custom, auto-tunable Helion GPU kernels in SGLang, delivering **1.03–3.95× kernel-level speedups** on NVIDIA and AMD GPUs and a 3% end-to-end speedup on AMD MI350. Meta says the model has additional safety post-training and met the thresholds required by its Meta Superintelligence Scaling Framework; detailed results are in the model card, with a full evaluation report promised soon. The weights are released under the Muse Model License with limited commercial terms, which Meta says allow many developers to use the model for free. Meta does not list benchmark scores, pointing instead to an evaluation methodology report. ## The bits that matter - Meta says Muse Spark Open has 816B total parameters and 42B active parameters, with a 1M-token context window. It is natively multimodal and optimized for coding, long-horizon agentic work, complex software engineering, and multi-step reasoning. - Meta says the model runs on two Apple M5 Ultra Mac Studios using MLX, with each chip able to address up to 512 GB of unified memory and 1.2 TB/s of memory bandwidth. The company claims the pair behaves like one supercomputer with a terabyte of unified memory. - The release is available through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI, with weights on Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models. Inference engines include SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth. Meta points developers to PyTorch's TorchTitan for further tuning. - Custom Helion GPU kernels in SGLang deliver 1.03–3.95× kernel-level speedups on NVIDIA and AMD GPUs and a 3% end-to-end speedup on AMD MI350, according to Meta. Meta also reports on-device agentic tests with Muse Code, including iOS changes to Ice Cubes and end-to-end DaVinci Resolve editing. - Meta says Muse Spark Open received additional safety post-training and met the thresholds of its Meta Superintelligence Scaling Framework. The weights use the Muse Model License with limited commercial terms; Meta promises a full evaluation report and does not include benchmark scores in the release notes. ## Discussion themes Analyzed: 2026\-10\-16T06:07:03\.693416\+00:00 Analysis sample: Based on 25 of 41 usable stored comments. Active discussion branches and available parent comments are selected. The analysis input was further shortened to fit its context limit. This sample may omit parts of the full thread. Selected themes do not measure community opinion or how common a view is. ### License revenue trigger tests open\-weight label The Muse Model License permits internal use, research, and most commercial product integration, but a model-as-a-service provider whose aggregate gross revenue exceeds $20 million over twelve months must notify Meta and sign a separate agreement. That carve-out drew arguments that the release is not open source, while a counterpoint notes Meta and its defenders never claimed OSI-approved open source. The threshold is described as not especially harsh, though it still conditions a core freedom. Sources: [Comment 50108164]() · [Comment 50108238]() · [Comment 50108900]() · [Comment 50108948]() ### Copyrightability of weights remains contested A legal argument holds that model weights are excluded from copyright because they result from automated optimization rather than human authorship, citing Thaler v. Perlmutter and the Copyright Office compendium. A counterpoint distinguishes the weights from model outputs: the weights themselves are prima facie human-authored artifacts and can carry Meta's license, while outputs may lack copyright coverage. A further position states flatly that weights are facts and facts are not copyrightable. Sources: [Comment 50108559]() · [Comment 50108715]() · [Comment 50108875]() ### Benchmark gap and reasoning\-level blind spot Third-party comparisons place Muse Spark Open below Qwen3.8 Flash Next and GLM 5.3 Flash, both of which can run on a single B300, suggesting Meta trails in open-weight performance. The comparison is limited because Artificial Analysis reports at the highest reasoning level, which may not reflect high-volume production use. Low and medium settings are requested, though the label 'medium' is warned to be arbitrary and gameable. Sources: [Comment 50108293]() · [Comment 50108665]() ### Consumer AI moat and platform control The claim that Meta will win consumer AI because Muse is sticky drew challenges: Muse has not existed long enough to establish stickiness, and the interface may be OpenClaw wrapped in Meta branding that any company could replicate. A counterargument holds that Google, Apple, and Microsoft control the underlying platforms and can restrict permissions or data handling, while Meta's hardware responses have been a failed phone and a speculative headset. A separate dispute over whether the rent market for intelligence has failed cites booming Anthropic revenue against claims of unprofitability. Sources: [Comment 50107247]() · [Comment 50107306]() · [Comment 50107351]() · [Comment 50107505]() · [Comment 50107537]() · [Comment 50107594]() · [Comment 50107610]() · [Comment 50107685]() · [Comment 50107702]() · [Comment 50107763]() · [Comment 50107960]() · [Comment 50107985]() ### Cost per task and daily\-driver preferences A cost-per-task comparison puts GPT Sol 6.1 high at $0.32 and Mimo 2.6 pro at $0.13 against GLM Max at $2.01, questioning why the latter would be a daily driver. A separate report describes using GLM 5.3 Flash via Ollama Cloud as a capable daily driver and rates Muse Spark 1.3 higher, though that is a self-reported preference rather than a measured benchmark. Sources: [Comment 50110035]() · [Comment 50109534]() ### Deployment breadth and hardware bar The release is packaged for many inference engines, from Unsloth to MLX, which lowers integration friction for developers. More detail is requested on what that means in practice. The hardware comparison remains a constraint: competing open-weight models cited in the benchmark thread can run on a single B300, while Muse Spark Open's desktop demo uses two Apple M5 Ultra Mac Studios. Sources: [Comment 50107063]() · [Comment 50107941]() · [Comment 50108293]() ## Sources & coverage AI-generated summary · 2026\-10\-16T06:07:03\.831583\+00:00 Based on 25 of 41 usable stored comments, selected by depth and branch activity. This is a sample of the discussion. The model input was further shortened to fit its context limit. Article text may also be shortened. Generated using deepseek\-ai/DeepSeek\-V4\.1\-Flash. Check the linked sources for full context.