Today for AI
HOT RADAR
llmHEAT 0.0°

AI CURATED HOT STORY · 10/15/2026

Meta Unveils Muse Spark Open: 816B MoE Model Enables Million-Token Context and Dual Mac Studio Local Inference

19 reports archived0 independent sourcesupdated 10/20/2026, 23:20:00
Synthesis & Latest Updates

Meta has officially open-sourced Muse Spark Open, its most powerful US-based open-weight multimodal model to date. Utilizing a Mixture-of-Experts (MoE) architecture with 816 billion total parameters and 42 billion active parameters, the model supports native text, image, audio, and video inputs. It features a massive 1 million token context window, specifically optimized for long-horizon agentic coding tasks and complex software engineering workflows.

Cross-verified reports indicate that Muse Spark Open performs comparably to top Chinese open-source models like Qwen and GLM in benchmarks. A key breakthrough is its ability to run efficiently on two M5 Ultra Mac Studios via collaboration with Apple and optimization through the MLX framework. This marks a significant advancement in local inference capabilities for high-end consumer hardware, despite requiring substantial memory resources (at least 700GB VRAM or dual 512GB RAM setups).

The release signifies Meta's strong return to the open-source race. The license permits free commercial use and fine-tuning but imposes restrictions on API resellers with annual revenues exceeding $20 million. This strategy balances ecosystem growth with commercial protection while pushing the boundaries of local AI deployment on Apple Silicon.

LATEST/Muse Spark Open is now live on Arena benchmarks and OpenRouter, priced at $1.25 per million input tokens and $4.25 per million output tokens. The community is actively evaluating its performance on dual M5 Ultra Mac Studio clusters and exploring fine-tuning strategies for specific codebases.

Coverage timeline

Total 19 reports · Latest first
  2. InfoQ 中文T2·85 pts

    Meta Opens Sources 816B-Param Muse Spark for Agents, Lags in Coding Benchmarks

    • Muse Spark Open features 816B total parameters with ~42B active per token and supports a 1M token context window.
    • Core strengths lie in end-to-end tool invocation and failure recovery for long-horizon agents, but it lacks native image/audio/video generation capabilities.
    • Local deployment requires two M5 Ultra Mac Studios; it scores lower than Zhipu and Kimi on coding benchmarks like Terminal-Bench.
  3. Interconnects (Nathan Lambert)T1·78 pts

    Meta Releases Muse Spark Open, Closing Gap with Chinese Frontier Models

    • Meta's Muse Spark Open is confirmed as the strongest US-made open-weight model, raising the bar for Western open source.
    • The performance gap between top US and Chinese open models has narrowed by roughly half, shifting from irrelevant to relevant.
    • Beyond Meta, Reflection's Beam and Mistral Large 4 demonstrate strong capabilities, signaling more high-quality US open models are coming.
  4. Latent SpaceT1·68 pts

    Meta Releases Muse Spark Open: High-Performance Multimodal MoE Model

    • Muse Spark Open sets a new record for US open-weight models with a score of 39, significantly outperforming NVIDIA's Nemotron 3.5 Super VL.
    • NVFP4 expert-only quantization reduces 1.6TB BF16 weights to ~530GB, lowering barriers for high-end consumer hardware deployment.
    • The model shows strong multimodal understanding and general intelligence but ranks low in agentic evaluations, indicating capability imbalances.
  5. Artificial Analysis (@ArtificialAnlys)T1·88 pts

    Meta Releases Muse Spark Open: 816B MoE Model Tops US Open-Weights Leaderboard with Multimodal Input

    • Muse Spark Open tops the US open-weights leaderboard with a score of 39, outperforming Meta's previous best 30B dense model by 22 points.
    • Utilizing an 816B total/42B active parameter MoE architecture, it achieves exceptional token efficiency, averaging only 40k output tokens per task.
    • Natively supports text, image, video, and audio inputs, showing strong performance on AutomationBench-AA with lower hallucination rates than predecessors.
  6. IT之家 · 智能时代T2·78 pts

    Apple and Meta Partner to Run Muse Spark Open Locally on Dual M5 Ultra Mac Studios

    • Meta and Apple break years of commercial and regulatory rivalry to collaborate publicly on local open-source model execution.
    • Muse Spark Open features 816B total parameters (42B active) and a 1M token context window, optimized for coding and agent tasks.
    • The demo runs on two unreleased 512GB M5 Ultra Mac Studios using the MLX framework for efficient local inference.
  8. IT之家 · 智能时代T2·92 pts

    Meta Releases Muse Spark Open: 816B MoE Model with 1M Context Runs Locally on Dual M5 Ultra Macs

    • Muse Spark Open features 816B total parameters with 42B active, supporting a 1M-token context window.
    • The model is specifically optimized for code generation, agentic workflows, and long-horizon complex tasks.
    • Enabled by the MLX framework, it is the first in its class to run locally on two M5 Ultra Mac Studios.
    • Weights are publicly available on platforms like Hugging Face and compatible with major inference engines such as vLLM and SGLang.
  9. Rohan Paul (@rohanpaul_ai)T2·88 pts

    Meta and Apple Collaborate: Muse Spark Open 816B MoE Model Runs Locally on Dual Mac Studios via MLX

    • Meta releases Muse Spark Open, an 816B parameter open-weight model designed for coding agents, scoring 89.2 on MCP-Atlas tool-use tests.
    • Apple optimizes the release for MLX, utilizing MoE sparse activation to allow two linked 512GB Mac Studios to host the full model efficiently.
    • The NVFP4 quantized version requires at least 700GB GPU memory; since each M5 Ultra offers 512GB, a dual-machine cluster is necessary to manage capacity and bandwidth.
  10. Rohan Paul (@rohanpaul_ai)T2·85 pts

    Meta Releases Muse Spark Open: 816B MoE Model Enables Local Execution on Mac Studio with 42B Speed

    • Muse Spark Open is Meta's first open-weight frontier model, based on Muse Spark 1.3 with added safety post-training against misuse.
    • Utilizes an extremely sparse MoE architecture (816B total / 42B active parameters) to balance knowledge capacity with inference latency.
    • Achieves local quantized execution on Mac Studio via Apple collaboration, leveraging M5 Ultra's high-bandwidth unified memory (approx. 344GB VRAM for 3-bit quant).
  12. The DecoderT2·85 pts

    Meta Releases Muse Spark Open: 816B Multimodal Model with 1M Context Rivals Chinese Giants

    • Muse Spark Open features 816 billion parameters and supports multimodal inputs including text, image, video, and audio.
    • It offers a 1-million-token context window and matches the benchmark performance of leading Chinese models like Qwen and GLM.
    • The model is explicitly optimized for local inference on Apple's M5 Ultra hardware.
  13. Arena.ai (@arena)T1·88 pts

    Meta Releases Muse Spark Open: An 816B MoE Model with 1M Context for Agentic Tasks

    • Muse Spark Open uses an 816B total / 42B active parameter MoE architecture to balance performance and inference cost.
    • Supports a 1 million token context window, enhancing capabilities for large codebases and long documents.
    • Features native multimodality across text, image, video, and audio, designed specifically for long-horizon agentic workflows.
  14. The Verge AIT2·68 pts

    Meta and Apple Optimize Muse Spark Open for Dual M5 Ultra Mac Studios via MLX

    • Meta launches Muse Spark Open, positioned as the most capable US open-weight model currently available.
    • Apple and Meta engage in rare public collaboration, optimizing model performance on Apple Silicon via the MLX framework.
    • The model supports local inference on a cluster of two M5 Ultra Mac Studios, overcoming desktop compute bottlenecks.
  15. OpenRouter (@OpenRouter)T1·86 pts

    Meta Releases Muse Spark Open: 816B MoE Model for Agentic Coding with 1M Context and Multimodal Input

    • Positioning: Claimed as the strongest open-weight model from a US lab, optimized for long-horizon tasks and complex software engineering.
    • Architecture: 816B total parameters with 42B active (MoE), balancing performance and inference cost.
    • Capabilities: Supports 1M context window and fully multimodal input including text, image, video, audio, and PDF.
  16. AWS Machine Learning BlogT1·38 pts

    Meta's Muse Spark Open Coding Model Now Available on Amazon SageMaker JumpStart

    • Muse Spark Open is Meta's open-weight model optimized for agentic coding workflows, focusing on first-attempt accuracy and tool calling.
    • The model supports a 1 million token context window, suitable for large codebases and multi-file refactor tasks.
    • Two variants are available on SageMaker JumpStart: BF16 (`Muse-Spark-Open-816B-A42B`) and NVFP4 quantized, allowing users to balance performance and cost.
  17. 宝玉 (@dotey)T2·88 pts

    Meta Releases Muse Spark Open: 816B MoE Model with Commercial License Requires Dual Mac Studios for Local Inference

    • Muse Spark Open is Meta's first open-weight large-scale Muse model, optimized for coding and long-horizon agentic tasks.
    • The 816B total/42B active parameter MoE architecture reduces compute but demands massive memory for full loading.
    • The license permits commercial use, modification, and distillation, but requires separate agreements for API resellers with >$20M annual revenue.
  19. vLLM 官方博客T1·68 pts

    vLLM Adds Day 0 Support for Meta's 816B MoE Muse Spark Open Model

    • Muse Spark Open is an 816B parameter sparse MoE model with 42B active parameters and a 1M context window, distilled from Muse Glimmer 30B.
    • vLLM enables Day 0 support with comprehensive parallelism (TP/PP/EP) on Nvidia, AMD, and TPU, plus DSpark speculative decoding for accelerated inference.
    • The model features native multimodal capabilities (image, video, audio) specifically engineered for long-horizon agentic coding workflows.