# Meta Muse Spark Open ![Organization](https://img.shields.io/badge/Org-Meta-1877f2) ![Context](https://img.shields.io/badge/Context-1M%20tokens-1f6feb) ![License](https://img.shields.io/badge/License-Muse%20Model%20License-d29922) ![Modality](https://img.shields.io/badge/Modality-multimodal-8957e5) ![Released](https://img.shields.io/badge/Released-Oct%202026-555) > **Quick answer:** Meta Muse Spark Open is Meta Superintelligence Labs' open-weights 816B-total/42B-active MoE model (released October 15, 2026), scoring 89.2% MCP-Atlas, 89.0% GPQA Diamond, and 80.4% OSWorld Verified with a 1M-token context window. Released under the custom Muse Model License (source-available, revenue-threshold clause for Model-as-a-Service). ## At a Glance **Where Muse Spark Open leads** - 89.2% MCP-Atlas tool-use accuracy, ahead of Nemotron 3 Ultra (63.1%), Beam (78.7%), and Inkling (76.0%) - Strong agentic/OSWorld computer-use score (80.4% OSWorld Verified) — notably higher than any other model in its own comparison table except Kimi K3 (84.8%) - 1M-token context, natively multimodal (text, image, video, audio input) **Where it lags** - Trails GLM 5.3 and Kimi K3 on GPQA Diamond (89.0% vs 91.7%/93.5%) and most coding sub-benchmarks (DeepSWE 1.1, SWE-Atlas-QnA, Terminal-Bench 2.1) - Lower HLE/Omniscience knowledge scores relative to Qwen3.8 and GLM 5.3 on some cuts despite leading on AA Omniscience **Best for:** teams that need open weights with strong agentic tool-use (MCP/OSWorld) rather than pure leaderboard-topping knowledge or coding scores. ## What Muse Spark Open Is Muse Spark Open is Meta Superintelligence Labs' open-weights release of its Muse Spark model family, an 816B-total/42B-active-parameter Mixture-of-Experts transformer (256 experts, top-8 routing) with a dedicated 1.8B-parameter perception encoder for image/video/audio input. It targets multi-step agentic reasoning, tool use, and coding-agent workflows, and ships with additional safety post-training intended for standalone deployment without an external guardrail layer. It is distributed under the "Muse Model License" — permissive for most uses (modify, fine-tune, distill, redistribute) but requiring attribution ("Built with Muse Spark Open") and a separate agreement with Meta if a Model-as-a-Service deployment built on it exceeds $20M/year in aggregate revenue. Weights are available in BF16 and NVFP4; NVFP4 needs ~700GB VRAM, BF16 ~2TB. Supported runtimes include SGLang, vLLM, Hugging Face Transformers, TensorRT-LLM, Unsloth, and llama.cpp (with an MLX path for dual Mac Studio M5 Ultra setups). ## Specifications | Field | Value | |---|---| | Organization | Meta | | Parameters | 816B total / 42B active (MoE, 256 experts, top-8 routing) | | Context window | 1,000,000 tokens | | Architecture | Mixture-of-Experts Transformer + 1.8B-param Perception Encoder | | License | Muse Model License (custom, source-available with MaaS revenue clause) | | Release date | 2026-10-15 | | Modality | multimodal (text, image, video, audio in / text out) | ## Pricing Open weights — free to download from Hugging Face (`meta-models/Muse-Spark-Open-816B-A42B`, gated access). No Meta-hosted API pricing published at release; third-party inference providers host it independently. ## Context Window Muse Spark Open has a **1,000,000-token context window** at inference — roughly 1,500+ pages of text in a single request. Combined with its agentic training, this supports long-horizon multi-step tool-use workflows and large-codebase or long-document analysis. ## Public Benchmark Scores | Benchmark | Score | Source | Date | |---|---|---|---| | [MCP-Atlas](/benchmarks/scale-ai/mcp-atlas) | 89.2 | [Meta model card](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B) | 2026-10 | | [GPQA Diamond](/benchmarks/rein-et-al/gpqa-diamond) | 89.0 | [Meta model card](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B) | 2026-10 | | [OSWorld Verified](/benchmarks/osworld/osworld-verified) | 80.4 | [Meta model card](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B) | 2026-10 | | [Toolathlon-Verified](/benchmarks/hkust-nlp/toolathlon-verified) | 72.8 | [Meta model card](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B) | 2026-10 | | [JobBench](/benchmarks/jobbench) | 61.8 | [Meta model card](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B) | 2026-10 | | [SciCode](/benchmarks/scicode) | 58.5 | [Meta model card](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B) | 2026-10 | | [DeepSWE](/benchmarks/datacurve/deepswe) | 60 | [Meta model card](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B) | 2026-10 | | [SWE Atlas – Codebase Q&A](/benchmarks/raghavendra-et-al-scale-ai/swe-atlas-codebase-qa) | 58 | [Meta model card](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B) | 2026-10 | | [ZeroBench](/benchmarks/zerobench) | 50.0 | [Meta model card](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B) | 2026-10 | | [BabyVision](/benchmarks/babyvision) | 80.2 | [Meta model card](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B) | 2026-10 | | [GDPVal-AA v2](/benchmarks/gdpval-aa-v2) | 1401 | [Meta model card](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B) | 2026-10 | | [AA Omniscience](/benchmarks/aa-omniscience) | 25 | [Meta model card](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B) | 2026-10 | | [Artificial Analysis Intelligence Index](/benchmarks/aa-intelligence-index) | 39 | [Meta model card](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B) | 2026-10 | *Scores above are reported by Meta's own model card and shown for context. They are not Benchgen measurements. Figures use xhigh reasoning effort, temperature=1.0, top_p=1.0, top_k=disabled — see the source for full methodology.* **Excluded from the table above (bucket-C, documented per scout procedure):** - **HLE-Diamond (no tools) — 26.1** and **HLE (w/ tools) — 61.0**: the vendor table reports two distinct HLE variants/conditions that do not clearly map to our existing generic `humanitys-last-exam` entity's methodology (tool-use condition and "Diamond" subset unconfirmed against our entity's task set) — excluded rather than guessed. - **Terminal-Bench 2.1 (Terminus 2 agent) — 76.4**: our local entities are Terminal-Bench 3/4/Hard; no 2.1 entity exists and the version doesn't match an existing methodology — excluded. - **SimpleVQA (w/ tools) — 73.6**: no matching local benchmark entity (checked `docvqa`, `infovqa`, `mtvqa`, `pmc-vqa`, `worldvqa-forceanswer` — none match); insufficient public methodology/task-count detail surfaced in this run to justify a new bucket-B config — excluded, flagged for follow-up. ## Related Models & Benchmarks - Compare: [Kimi K3](/models/moonshot/kimi-k3) · [GLM 5.3](/models/zhipu/glm-5-3) - Benchmarks: [MCP-Atlas](/benchmarks/scale-ai/mcp-atlas) · [GPQA Diamond](/benchmarks/rein-et-al/gpqa-diamond) · [OSWorld Verified](/benchmarks/osworld/osworld-verified) ## Frequently Asked Questions

What is Meta Muse Spark Open?

It's Meta Superintelligence Labs' open-weights release of its Muse Spark model family — an 816B-total/42B-active MoE model for agentic, multimodal, and coding workloads, released October 15, 2026.

What is Muse Spark Open's context window?

1,000,000 tokens — roughly 1,500+ pages of text in one request.

How much does Muse Spark Open cost?

Open weights — free to download from Hugging Face. Self-hosting requires ~700GB VRAM (NVFP4) or ~2TB (BF16); no first-party Meta API pricing published at release.

Is Muse Spark Open open source?

It's released under the custom "Muse Model License" — source-available and permissive for most uses, with an attribution requirement and a separate-agreement clause for Model-as-a-Service deployments exceeding $20M/year revenue.

What is Muse Spark Open's knowledge cutoff?

Not disclosed in the public model card as of this writing.

--- *Specs and scores sourced from Meta's official Hugging Face model card. Last updated 2026-10-17.*

Muse Spark Open 816B A42B — Evaluations

Evaluations