Muse Spark Open: From Open Release to Local Runs

Episode 1 · Meta Open-Sources Muse Spark Open, an 816B-Parameter Open-Weight Model (2026-10-16, 64 posts)

Meta released and open-weighted Muse Spark Open on October 16, with the company and Chief AI Officer Alexandr Wang calling it "the most capable American open-weights model to date." The model targets coding and agentic use cases, excelling at long-horizon tasks, complex software engineering, and multi-step reasoning. Weights are available on Hugging Face under the Muse Model License.

Confirmed

• Core specs: 816B total parameters, 42B activated (MoE architecture), 1M token context window, natively multimodal (text/image/video/audio)
• Muse-Spark-Open-816B-A42B under the meta-models org on Hugging Face ships in BF16 and NVFP4; unsloth also has it available
• Day-0 support in transformers and llama.cpp; the community already has an MLX 3bit quantized version; config files reveal low/medium/high/xhigh reasoning effort tiers
• Zuckerberg announced the open release and thanked Apple for the collaboration, saying it runs well on the new Mac Studio; Hugging Face CEO Clement Delangue reshared a demo running GGUF Q2K quantization across two Mac Studios with a single llama serve command
• Alexandr Wang said the model is available immediately via seven API platforms—Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI—with support for mainstream inference engines like vLLM/SGLang
• Official benchmark tables show the model surpassing GLM and Kimi on multiple leaderboards; it has undergone additional safety post-training

Unconfirmed

• @op7418's post claims 1T token context, contradicting the 1M token figure from official sources and multiple others; the official 1M stands

Why it matters

This is another major move in Meta's open-source strategy and a direct declaration of "America's strongest open-weights model." The combination of multimodality, ultra-long context, and agentic capability—plus full-stack deployment support from local Macs to cloud APIs—makes it appealing to both developers and the local-deployment community, and @bindureddy sees it as a key step for Meta to catch up with other AI labs.

• Meta's Muse Spark Open: full specs, Helion kernel speedups and benchmark tables — AIatMeta · 2026-10-16
• Meta's Muse Spark Open: 816B/A42B, 1M context, top US open-weight model — alexandr_wang · 2026-10-16
• Zuckerberg: Meta's Muse Spark Open weights out today, runs on Mac Studios — AIatMeta · 2026-10-16
• Meta's Muse Spark Open (816B/A42B) weights land on Hugging Face — Bob_SUS · 2026-10-16
• Unsloth Hosts Meta's Muse-Spark-Open-816B-A42B on Hugging Face — Bob_SUS · 2026-10-16
• Scale AI's open muse spark model hits seven API platforms, runs on vLLM, SGLang, MLX — alexandr_wang · 2026-10-16
• Meta releases Muse Spark Open: 816B/42B active, 1M context, most capable US open-weight model — AIatMeta · 2026-10-16
• Meta releases Muse Spark Open: 816B/42B MoE, 1M context, its most capable open model yet — AIatMeta · 2026-10-16
• Meta releases Muse Spark Open: most capable US open-weight model, runs on two Mac Studios — AIatMeta · 2026-10-16
• Meta releases Muse Spark Open: 816B/42B active, 1M context, natively multimodal — MaziyarPanahi · 2026-10-16
• Meta quietly open-sources Muse Spark, an 816B-parameter MoE model with 42B active params — victormustar · 2026-10-16
• Alexandr Wang announces muse spark open weights, "best US open-source model," runs on Mac Studio — LukeZettlemoyer · 2026-10-16
• Meta officially launches Muse Spark Open: 816B total, 42B active, 1M context — airesearch12 · 2026-10-16
• Meta open-sources Muse Spark Open: 816B MoE with 42B active, 1M context, strong benchmarks — testingcatalog · 2026-10-16
• Meta open-sources Muse Spark, first base model from its new MSL org — RylanSchaeffer · 2026-10-16
• Meta releases 816B-A42B open-weight Muse Spark Open; Apple MLX optimized for it — BenBajarin · 2026-10-16
• Muse Spark Open: Meta's 816B multimodal MoE built for agentic coding hits Hugging Face — mervenoyann · 2026-10-16
• Meta quietly releases Muse Spark Open: 816B-param MoE with 42B active, on Hugging Face — victormustar · 2026-10-16
• Meta open-sources Muse Spark Open: 816B MoE that runs on two Mac Studios — op7418 · 2026-10-16
• Meta releases Muse Spark Open: 816B open-weight model that runs on two M5 Ultra Mac Studios — op7418 · 2026-10-16

44 more related posts →

Episode 2 · Meta Open-Sources 816B MoE Model Muse Spark Open; Baseten Offers Day 0 API (2026-10-16, 2 posts)

Meta released Muse Spark Open, an open-weights 816B MoE model with 42B active parameters, 1M context, and omni-modal inputs; Baseten launched its API on day 0 at $1.25 per million input tokens in NVFP4 precision.

• Baseten serves Meta's Muse Spark Open on day 0 at $1.25/M input tokens — baseten · 2026-10-16
• Meta releases Muse Spark Open: 816B MoE, 42B active, 1M context, day-0 on Baseten — baseten · 2026-10-16

Episode 3 · vLLM and Inferact Optimize Muse Spark Open Serving (2026-10-16, 2 posts)

vLLM detailed cross-hardware optimizations for Muse Spark Open, reaching 960 tokens/s on TPU, while Inferact trained the DSpark speculative decoding draft model that proposes up to 7 draft tokens per step.

• DSpark draft model for Muse Spark Open proposes 7 tokens per pass with 23-42% acceptance — vllm_project · 2026-10-16
• vLLM details Muse Spark Open cross-hardware stack: 960 tokens/sec on 8 TPU v7 chips — vllm_project · 2026-10-16

Episode 4 · Apple Runs Near-Trillion-Parameter Model on 4 Mac Studios at 36 Tokens/s (2026-10-16, 3 posts)

Apple's MLX team demonstrated running the near-trillion-parameter open model Muse Spark Open (816B) locally on a cluster of four M5 Ultra Mac Studios using MLX and Thunderbolt-based RDMA, achieving over 36 tokens per second in decoding.

• Apple runs near-trillion-parameter Muse Spark Open on 4 M5 Ultra Macs at 36.7 tok/s — EAccelerate_42 · 2026-10-16
• Apple MLX Runs 816B Muse Spark Open at 36+ Tokens/s on Four Mac Studios — andrejusb · 2026-10-16
• Apple: four M5 Ultra Mac Studios run near-trillion-param Muse Spark Open at 36 tok/s via MLX and RDMA — rxwei · 2026-10-16