Meta Muse Spark Open 816B Model Released for Coding and Agentic Tasks

Meta released Muse Spark Open on October 15, 2026. The open-weight Mixture-of-Experts model contains 816 billion total parameters with 42 billion active per token and supports a 1-million-token context window. It targets coding, multi-step reasoning, tool use, and long-horizon agentic workflows.

Technical Specifications and Capabilities

The architecture uses a sparse MoE transformer with perception encoders for text, image, audio, and video inputs while producing text output. Each layer contains 256 experts and activates eight per token. Meta provides checkpoints in BF16 format at approximately 1.65 TB and in NVFP4 format at 537 GB. Training covers data from more than 100 languages.

Developers can run the model locally on two Apple M5 Ultra Mac Studios through the MLX framework, which requires roughly 1 TB of unified memory. On GPUs, the NVFP4 version fits on four Blackwell or eight H200 accelerators, while the BF16 version needs about 2 TB of VRAM. Supported inference engines include SGLang and vLLM with optimizations for NVIDIA, AMD, Intel, and Apple hardware.

Weights are available on Hugging Face under the Muse Model License and through Amazon SageMaker, Google Cloud Model Garden, and Microsoft Foundry. Hosted inference runs on platforms such as Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter, and Together AI at roughly $1.25 per million input tokens and $4.25 per million output tokens.

Performance Benchmarks and Industry Context

Meta-reported scores show strong results on agentic and tool-use tasks. The model reaches 89.2 on MCP-Atlas, 61.8 on JobBench, 76.4 on Terminal-Bench 2.1, and 60 on DeepSWE 1.1. It also records 39 on the Artificial Analysis Intelligence Index. Performance leads listed rivals in tool-use categories while remaining competitive on pure coding benchmarks against certain Chinese models.

Muse Spark Open builds on the earlier proprietary Muse Spark series that debuted in April 2026 and on the smaller Muse Glimmer model released in August. The release continues Meta’s pattern of making advanced models available in open weights while retaining its most capable systems behind closed APIs. Additional safety post-training allows standalone deployment, though Meta recommends adding further guardrails for production use.

Community adoption is expected to focus on fine-tuning through TorchTitan and PyTorch for specialized coding agents and enterprise workflows. Organizations evaluating the model can compare its agentic scores directly against existing open-weight alternatives before committing infrastructure resources.

Deployment Options and Next Steps

Enterprises can download the model for on-premises inference or integrate it through the listed cloud providers. Pricing on hosted platforms supports both experimentation and production-scale agent deployments. Meta has published detailed model cards and evaluation results to help teams assess fit for specific coding or reasoning workloads.

Teams planning local deployment should verify hardware compatibility with the NVFP4 checkpoint to reduce memory requirements. Further optimizations from the open-source community are anticipated in the coming months, particularly around inference speed on consumer-grade and enterprise GPU clusters.