Meta opens Muse Spark: 816B weights, a scoreboard that admits Kimi wins coding, and two Mac Studios as the demo
Two days after shipping Muse Wonder as a closed subscription, Meta released the weights of Muse Spark Open on October 15 and called it the most capable US open-weight model to date. The wording is careful. Not the most capable open model, the most capable American one. The model is an 816B-parameter mixture of experts with 42B active per token, a 1M-token context, native text, image, video and audio input, and it is tuned for coding and agent work. Weights are gated on Hugging Face under the Muse Model License, which Meta describes as having limited commercial terms. That is not open source by any license body's definition, and the Hugging Face repo requires a login and approval. Meta's own account posted the launch and it passed 1.06 million impressions with 11,000 likes and 659 quote posts. The HN thread hit 899 points.
Meta published the comparison table as an image, so here it is in text. On MCP-Atlas, the 1,000-task Model Context Protocol benchmark, Muse Spark Open scores 89.2 against Qwen3.8 at 84.5, GLM 5.3 at 84.2, Kimi K3 at 82.3 and Beam at 78.7. On JobBench it posts 61.8 against GLM 5.3 at 61.4 and Inkling at 31.1. On HLE-Diamond without tools it leads at 26.1. Then the coding rows flip. Terminal-Bench 2.1 is 76.4 for Muse against 85.0 for Kimi K3, 83.9 for GLM 5.3, 82.0 for Qwen3.8 and 80.1 for Beam. DeepSWE 1.1 is 60 for Muse against 68 for Kimi K3 and 67 for GLM 5.3, and that row is labelled best available or self-reported harness, so each model ran in its own vendor harness. GDPval-AA v2 is 1401 for Muse against 1653 for GLM 5.3. The methodology PDF says comparators ran at xhigh or max effort and that Beam and Mistral Large 4 numbers were lifted from their launch posts because weights were not available. A lab publishing a table where it loses most of the coding rows is still unusual enough to note.
The launch demo is the actual message. Meta says Muse Spark Open is the first model of its class that runs on a desktop, meaning two Apple M5 Ultra Mac Studios with 512GB of unified memory each, linked so they behave like one machine with a terabyte of memory, running through MLX. Apple published its own post the same day showing a four-machine M5 Ultra cluster over Thunderbolt RDMA decoding at more than 36 tokens per second at 8K context. Meta's on-device demos were agentic by design: Muse Code shipping real changes to the Ice Cubes iOS app, build included, and the model driving DaVinci Resolve end to end, reviewing footage frame by frame to pick and crop moments. The catch is that the 512GB Mac Studio does not ship until the end of October, so right now the local story is a cluster of four 256GB machines, which is a forty thousand dollar desktop.
Everything else about the release is distribution. Day-zero API access on Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter and Together. Weights on SageMaker, Model Garden, Hugging Face and Microsoft Foundry. Support in SGLang, vLLM, TensorRT-LLM, Transformers, MLX, llama.cpp and Unsloth, with Meta's auto-tunable Helion kernels giving 1.03 to 3.95 times kernel-level speedups in SGLang. An mlx-community 4-bit quant and a ggml GGUF were up within hours. A torrent group announced it was seeding the NVFP4 checkpoint eight hours after release, which is what gated weights look like in practice.
Read it next to Monday's Wonder launch and the strategy is two boards at once. Wonder is the closed subscription agent. Spark Open is the kernel Meta wants enterprises to run inside their own walls, with the data never leaving. The top HN comment put it bluntly: once capability is good enough, price decides, and Meta is betting on becoming the default everyone optimizes infrastructure for. The thing to watch is whether independent runs confirm the MCP-Atlas and JobBench leads, because those are the two numbers that say agent work rather than coding, and they are the two Meta chose to lead with.
Links: https://research.meta.ai/blog/muse-spark-open , https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B , https://machinelearning.apple.com/research/mlx-distributed-inference , https://x.com/AIatMeta/status/2110779754274292212 , https://news.ycombinator.com/item?id=50107379
← Back to all articles
Meta published the comparison table as an image, so here it is in text. On MCP-Atlas, the 1,000-task Model Context Protocol benchmark, Muse Spark Open scores 89.2 against Qwen3.8 at 84.5, GLM 5.3 at 84.2, Kimi K3 at 82.3 and Beam at 78.7. On JobBench it posts 61.8 against GLM 5.3 at 61.4 and Inkling at 31.1. On HLE-Diamond without tools it leads at 26.1. Then the coding rows flip. Terminal-Bench 2.1 is 76.4 for Muse against 85.0 for Kimi K3, 83.9 for GLM 5.3, 82.0 for Qwen3.8 and 80.1 for Beam. DeepSWE 1.1 is 60 for Muse against 68 for Kimi K3 and 67 for GLM 5.3, and that row is labelled best available or self-reported harness, so each model ran in its own vendor harness. GDPval-AA v2 is 1401 for Muse against 1653 for GLM 5.3. The methodology PDF says comparators ran at xhigh or max effort and that Beam and Mistral Large 4 numbers were lifted from their launch posts because weights were not available. A lab publishing a table where it loses most of the coding rows is still unusual enough to note.
The launch demo is the actual message. Meta says Muse Spark Open is the first model of its class that runs on a desktop, meaning two Apple M5 Ultra Mac Studios with 512GB of unified memory each, linked so they behave like one machine with a terabyte of memory, running through MLX. Apple published its own post the same day showing a four-machine M5 Ultra cluster over Thunderbolt RDMA decoding at more than 36 tokens per second at 8K context. Meta's on-device demos were agentic by design: Muse Code shipping real changes to the Ice Cubes iOS app, build included, and the model driving DaVinci Resolve end to end, reviewing footage frame by frame to pick and crop moments. The catch is that the 512GB Mac Studio does not ship until the end of October, so right now the local story is a cluster of four 256GB machines, which is a forty thousand dollar desktop.
Everything else about the release is distribution. Day-zero API access on Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter and Together. Weights on SageMaker, Model Garden, Hugging Face and Microsoft Foundry. Support in SGLang, vLLM, TensorRT-LLM, Transformers, MLX, llama.cpp and Unsloth, with Meta's auto-tunable Helion kernels giving 1.03 to 3.95 times kernel-level speedups in SGLang. An mlx-community 4-bit quant and a ggml GGUF were up within hours. A torrent group announced it was seeding the NVFP4 checkpoint eight hours after release, which is what gated weights look like in practice.
Read it next to Monday's Wonder launch and the strategy is two boards at once. Wonder is the closed subscription agent. Spark Open is the kernel Meta wants enterprises to run inside their own walls, with the data never leaving. The top HN comment put it bluntly: once capability is good enough, price decides, and Meta is betting on becoming the default everyone optimizes infrastructure for. The thing to watch is whether independent runs confirm the MCP-Atlas and JobBench leads, because those are the two numbers that say agent work rather than coding, and they are the two Meta chose to lead with.
Links: https://research.meta.ai/blog/muse-spark-open , https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B , https://machinelearning.apple.com/research/mlx-distributed-inference , https://x.com/AIatMeta/status/2110779754274292212 , https://news.ycombinator.com/item?id=50107379
Comments