Overview
Muse Spark Open is an open-weight model from Meta Superintelligence Labs, released on October 15, 2026 for coding and agentic work: long-horizon tasks, complex software engineering and multi-step reasoning. Meta calls it the most capable US open-weight model to date. It is a mixture-of-experts transformer with 816B total parameters, of which 42B are active per token (256 experts, 8 activated), and a context window of 1M tokens.
The model is natively multimodal. It takes text, image, video and audio input and returns text, and its training data spans more than 100 languages. Reasoning effort is selectable at low, medium, high or xhigh, with xhigh as the default; Meta recommends temperature 1.0 and top_p 1.0. Because the weights are public, Meta applied additional safety post-training and evaluated the model under its Superintelligence Scaling Framework before release.
Meta ships BF16 weights and an NVFP4 build. The model card asks for at least 700GB of GPU memory for NVFP4 or 2TB for BF16, and Meta says the model is optimized to run on two Apple Mac Studios with M5 Ultra chips using MLX, which together offer about a terabyte of unified memory. Meta tested that setup with Muse Code on agentic jobs such as shipping iOS changes to the Ice Cubes Mastodon client and driving DaVinci Resolve end to end.
| Released | 2026-10-15 |
|---|---|
| License | Muse Model License |
| Weights | Open weights |
| Parameters | 816B total · 42B active |
| Context | 1M |
| Architecture | Mixture-of-Experts transformer with a perception encoder: 256 experts, 8 activated per token. |
| Modalities | Text, Vision, Video, Audio |
| Status | Generally available (open weights on Hugging Face) |
Benchmarks
Muse Spark Open (xhigh) against the open-weight models Meta compared at launch. A dash means Meta reported no value. Higher is better except where marked ↓.
| Benchmark | Muse Spark Open | Inkling | Nemotron 3 Ultra | Beam | Mistral Large 4 | Qwen3.8-2.4T-A95B | GLM 5.3 | Kimi K3 |
|---|---|---|---|---|---|---|---|---|
| AA Intelligence Index | 39 | 26 | 23 | — | 38 | 40 | 45 | 44 |
| JobBench | 61.8 | 31.1 | 27.2 | — | — | 52.7 | 61.4 | 54.3 |
| MCP-Atlas | 89.2 | 76 | 63.1 | 78.7 | — | 84.5 | 84.2 | 82.3 |
| GDPVal-AA v2 | 1401 | 1079 | 1016 | — | 1424 | 1613 | 1653 | 1537 |
| OSWorld Verified | 80.4 | 40.1 | — | — | — | — | — | 84.8 |
| Toolathlon Verified | 72.8 | 45.5 | 34.3 | — | — | 72.5 | 73 | 76.5 |
| HLE-Diamond (no tools) | 26.1 | 11.2 | 8.9 | — | — | 17.3 | 16.4 | 22.2 |
| SciCode | 58.5 | 47 | 40.3 | 49.7 | 54.2 | 54.1 | 59 | 59.5 |
| HLE (with tools) | 61 | 46 | 37.4 | 51.2 | — | 56.2 | 62.5 | 56 |
| GPQA Diamond | 89 | 87.2 | 86.7 | 90.5 | — | 93.5 | 91.7 | 93.5 |
| DeepSWE 1.1 | 60 | — | — | 44 | 62 | 57 | 67 | 68 |
| Terminal-Bench 2.1 (Terminus 2) | 76.4 | 55.1 | 53.9 | 80.1 | — | 82 | 83.9 | 85 |
| StrongREJECT v2 (adaptive) ↓ | 0.5 | 86.4 | 80.7 | — | — | 65.6 | 62.9 | 66.6 |
This model's scores
- MCP-Atlas89.2
- OSWorld Verified80.4
- Terminal-Bench 2.1 (Terminus 2)76.4
- HLE (with tools)61
- GPQA Diamond89
Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.
Strengths
- Open weights in BF16 and NVFP4 for a 816B-total / 42B-active MoE, downloadable from Hugging Face, SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models
- Leads Meta's published comparison on tool use — MCP-Atlas 89.2 against 84.5 for Qwen3.8-2.4T-A95B and 82.3 for Kimi K3
- 1M-token context with text, image, video and audio input
- Runs on two M5 Ultra Mac Studios with MLX, as well as on GPU servers through SGLang, vLLM, TensorRT-LLM and Transformers
- Low harmful-compliance rate on Meta's safety table: StrongREJECT v2 of 0.5, where lower is better
- Hosted by Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI for teams without their own hardware
Best for
- Reach for it when you need a frontier-scale model you can self-host, for example inside a private network where data cannot leave.
- Reach for it for tool-heavy agents that call many MCP tools across long tasks, where it posts its best published scores.
- Reach for it to run a large model locally on two high-memory Mac Studios rather than a GPU server.
- Use Muse Glimmer instead when the model must fit one consumer GPU, or when you need an Apache-2.0 license rather than the Muse Model License.
How to access
| Provider | Model ID |
|---|---|
| Hugging Face (open weights) ↗ | meta-models/Muse-Spark-Open-816B-A42B |
| OpenRouter ↗ | meta/muse-spark-open |
FAQ
What hardware does Muse Spark Open need?
The Muse Spark Open model card asks for at least 700GB of GPU memory for the NVFP4 build or 2TB for BF16. Meta also says the model is optimized for two Apple Mac Studios with M5 Ultra chips using MLX, where each machine offers up to 512 GB of unified memory, so the pair holds the full model.
What license does Muse Spark Open use?
Muse Spark Open ships under the Muse Model License, which Meta describes as having limited commercial terms that still let many developers use the model for free. It is not Apache 2.0 — Meta's smaller open-weight model, Muse Glimmer, is the Apache-2.0 option in the Muse family.
How does Muse Spark Open compare to Kimi K3 and GLM 5.3?
In Meta's model-card table, Muse Spark Open scores 89.2 on MCP-Atlas, ahead of Kimi K3 (82.3) and GLM 5.3 (84.2), and 26.1 on HLE-Diamond without tools versus 22.2 and 16.4. It trails both on Terminal-Bench 2.1 (76.4 vs 85.0 and 83.9) and on the Artificial Analysis Intelligence Index (39 vs 44 and 45).
How do I serve Muse Spark Open?
The Muse Spark Open model card gives one-line launches for vLLM (vllm serve "meta-models/Muse-Spark-Open-816B-A42B"), SGLang (python3 -m sglang.launch_server with the same model path) and Docker Model Runner, plus Hugging Face Transformers loading through AutoProcessor. Meta also lists TensorRT-LLM, MLX, llama.cpp and Unsloth support.
How is Muse Spark Open different from Muse Spark?
Muse Spark and its 1.x updates are proprietary, closed-weight models served through the Meta Model API. Muse Spark Open publishes its weights: an 816B-total, 42B-active mixture-of-experts with a 1M-token context, which anyone can download and self-host under the Muse Model License.