Released in October 2026 with a 1M-token window, Muse Spark Open lets a team keep a tool-calling agent model on its own hardware in NVFP4 or full precision, with seven hosted APIs as a fallback. It suits agent builders who can fund a multi-GPU node and accept Meta's attribution rule, not open-ended assistants that need low refusal rates.
What changed
Muse Spark Open, released by Meta on October 15, 2026, scores 89.2 on MCP-Atlas in Meta's own evaluation at xhigh reasoning effort, the highest in its comparison table. It is an open-weight, natively multimodal mixture-of-experts model that accepts text, image, video and audio, built for coding and tool-using agents.
Where it sits
- 142 tok/stokens/sOutput speedHigher is better#18 / 53peer median 94 tok/scited: Artificial Analysis
- 89%% correctGPQA DiamondHigher is better#25 / 59peer median 87.4%per source, see benchmark scores
Ranks are against GA models on HokAI that publish the same figure; ties share a rank.
Provider: Meta AI · Family: Muse
Context window: 1,048,576 tokens
Input modalities: text, image, video, audio, tool-calls · Output: text, tool-calls
About Muse Spark Open
Meta published the weights of Muse Spark Open on October 15, 2026, on Hugging Face under the Muse Model License. It comes from the Superintelligence Labs team at Meta and uses a sparse mixture-of-experts design (a Transformer) with 816B total and 42B active parameters (256 experts, 8 routed per token), a 1.8B-parameter perception encoder and a 1M-token window at inference. Meta calls it the most capable US open-weight model to date, which is the company's own claim. Its lineage is described two ways: the PyTorch announcement says it shares a base model with Muse Spark 1.3, while Meta's developer page says it derives from the August Muse Spark release. The PyTorch post adds that this is the full-size model of the family, after the small dense Muse Glimmer released earlier. Either way it carries extra safety post-training meant to hold without system-level guardrails, and Meta says it met the release thresholds of its Superintelligence Scaling Framework.
On Meta's scorecard, run at xhigh reasoning effort, it posts 89.2 on MCP-Atlas, 80.4 on OSWorld Verified, 76.4 on Terminal-Bench 2.1 and 89.0 on GPQA Diamond, and Artificial Analysis gives it an Intelligence Index of 39. Against the open-weight rivals in that table it leads on MCP-Atlas (84.2 for GLM-5.3, 82.3 for Kimi K3) and on Humanity's Last Exam Diamond without tools (26.1 against 22.2 for Kimi K3), but it trails both on Terminal-Bench (83.9 and 85.0) and on the Intelligence Index (45 and 44). Read the table with care: Meta ran its own OSWorld and with-tools Humanity's Last Exam numbers, the third-party with-tools scores are self-reported, and GLM-5.3 and Kimi K3 were run at their maximum effort setting rather than xhigh.
Inputs can be text, images, video or audio, and the output is text. Reasoning effort is set per request to low, medium, high or xhigh, with xhigh the default and high or xhigh recommended for coding and agent work. Meta trained it to browse the web, to spawn and coordinate subagents, and to operate desktop, mobile and browser interfaces, and the Muse Code terminal agent can run on it. Audio input is a point to test yourself: the model card lists it as supported, whereas the HokAI record for Muse Spark 1.3 notes degraded audio understanding on that closed model.
Self-hosting is a cluster job. The model card asks for at least 700 GB of GPU memory for the NVFP4 checkpoint and 2 TB for BF16, and Meta's developer guide describes about 530 GB of NVFP4 weights that serve on four NVIDIA Blackwell Ultra GPUs or eight H200s, with BF16 at about 1.6 TB. Meta also shows the NVFP4 build running through MLX on two Apple Mac Studios with M5 Ultra chips and 512 GB of unified memory each, and the smallest Unsloth GGUF fits in roughly 300 GB of combined memory. SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, llama.cpp, MLX and Unsloth all support it, and the download sits behind a Hugging Face form that asks for a legal name, date of birth and organisation.
If you would rather not run GPUs, seven hosts served it at launch: Baseten, Fireworks AI, Groq (limited preview), Modal, Nebius, OpenRouter and Together AI. Amazon SageMaker JumpStart, Google Cloud Model Garden and Microsoft Foundry carry the weights for private deployments, and Hugging Face remains the download point and the place to find community quantizations. Artificial Analysis measured output speed from 100.9 tokens per second on Fireworks to 552.0 on Makora's NVFP4 endpoint in its provider table, so the host you pick matters nearly as much as the model.
Meta says it added safety post-training because open weights ship without a provider's guardrails: in its own tests with no system-level safeguards the model refused 99.2% of jailbreaks on an internal cyber benchmark, against 75.5% for Kimi K3 in the same setup. The trade-off shows in Meta's FORTRESS results, where it has the lowest adversarial risk score (5.8) in the table and also the highest refusal rate on benign prompts (17.4%).
Pick GLM-5.3 (from Z.ai) or Kimi K3 (from Moonshot AI) for terminal-heavy coding and a lower refusal rate, and Beam when the licence must be Apache 2.0. Other releases with downloadable weights sit in the open-weights model directory, and Meta's other models are listed on its provider page. Pick Muse Spark Open when tool-call accuracy, Meta's safety posture and a smaller footprint than the 2.8T-parameter Kimi K3 matter more.
Pricing
Meta's launch posts list no first-party price for Muse Spark Open. Downloading the weights carries no fee under the Muse Model License, so self-hosting costs GPU time rather than tokens. For hosted use, the [OpenRouter](/hub/tools/openrouter) API listing read on October 20, 2026 shows $1.25 per 1M input tokens, $4.25 per 1M output and $0.15 per 1M cached input. Artificial Analysis lists the same $1.25 and $4.25 as the lowest rates among Baseten, Nebius and Fireworks, with a blended price of $0.78 per 1M tokens on Baseten, Fireworks and Together AI against $1.55 on Nebius' FP4 endpoint. Rates differ by host and quantization, so confirm on the host's own page before budgeting. For comparison, the closed [Muse Spark 1.3](/hub/models/muse-spark) lists the same input and output rates on its Standard tier.
| Tier | Rate |
|---|---|
| Self-hosted (Muse Model License) | Free to download after accepting the licence on Hugging Face. Cost is GPU hardware. A separate Meta agreement applies to Model-as-a-Service sellers above US$20 million in revenue. |
| Hosted API (lowest-priced hosts) | OpenRouter API listing and Artificial Analysis provider table, read 2026-10-20. Rates vary by host and quantization. |
Key Features
- Sparse mixture of experts: Routes each token to 8 of 256 experts, so only a small slice of the total parameters runs per token, according to the model card.
- Million-token context: Runs at a 1M-token window at inference, and Meta's vLLM recipe sets the maximum model length to 1,048,576 tokens.
- Text, image, video and audio input: A 1.8B-parameter perception encoder handles visual input; the card lists audio as an input and text as the only output.
- Reasoning-effort control: Set reasoning_effort to low, medium, high or xhigh per request; xhigh is the default and any other value returns an error.
- Day-one runtime support: Launches with SGLang, vLLM, TensorRT-LLM, Transformers, llama.cpp, MLX and Unsloth, plus Helion kernels that Meta reports at 1.03 to 3.95 times faster at kernel level than hand-written kernels on NVIDIA GB300.
- Two Meta checkpoints: Full-precision reference weights plus an NVFP4 build at about a third of the size (about 530 GB), with AMD, Intel and Unsloth quantizations built for their own hardware.
Pros
- Leads Meta's open-weight scorecard on MCP-Atlas, the test closest to production agents that juggle many connected tools.
- The licence permits fine-tuning, distillation and redistribution, which API-only frontier models do not offer.
- Seven hosted APIs and three cloud catalogs at launch let a team trial it before buying any GPUs.
- Meta reports a 99.2% refusal rate on its internal cyber jailbreak test with no system-level safeguards, well above the 75.5% it measured for Kimi K3 under identical conditions.
Cons
- Refused 17.4% of benign FORTRESS prompts in Meta's own table, the highest of the six models with a reported figure.
- Trails GLM 5.3 and Kimi K3 on Terminal-Bench 2.1 (76.4 against 83.9 and 85.0) and on the Artificial Analysis Intelligence Index (39 against 45 and 44).
- Self-hosting needs a multi-GPU node, and the Hugging Face download is gated behind an identity form.
- The Muse Model License adds an attribution rule and a Meta agreement for large Model-as-a-Service sellers, which makes it open-weight but not open source.
Benchmarks
- MCP Atlas: 89.2% vendor-reported · 20 Oct 2026 — Tasks solved by calling external tools over the Model Context Protocol, % solved.
- GDPval-AA v2: 1,401 vendor-reported · 20 Oct 2026 — Real knowledge-work deliverables judged against professionals, run by Artificial Analysis.
- GPQA Diamond: 89% vendor-reported · 20 Oct 2026 — PhD-level science questions that are hard to search for, % correct.
- OSWorld Verified: 80.4% vendor-reported · 20 Oct 2026 — Tasks completed by operating a real desktop, % solved.
- Terminal-Bench 2.1: 76.4% vendor-reported · 20 Oct 2026 — Multi-step tasks completed in a real command line, % solved.
- AA Intelligence Index: 39 cited: Artificial Analysis · 20 Oct 2026 — Composite of 10 evaluations run by Artificial Analysis, 0 to 100.
- AA blended price: $0.78/M cited: Artificial Analysis · 20 Oct 2026 — Price per 1M tokens at a 3:1 input to output blend, as listed by Artificial Analysis.
- Output speed: 142 tok/s cited: Artificial Analysis · 20 Oct 2026 — Median tokens written per second as measured by Artificial Analysis.
A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.
Frequently Asked Questions
How much does Muse Spark Open cost in 2026?
Downloading the weights costs nothing, but Meta's model card calls for 700 GB or more of GPU memory in NVFP4 and 2 TB in BF16, so most teams will pay a host instead. The lowest rates tracked by Artificial Analysis are $1.25 per 1M input tokens and $4.25 per 1M output, and OpenRouter lists $0.15 per 1M for cached input. Artificial Analysis puts the average cost of running one Intelligence Index task at $0.84, measured on the xhigh setting that is the default.
What separates Muse Spark Open from Kimi K3?
On Meta's scorecard Muse Spark Open wins MCP-Atlas by about seven points, while Kimi K3 wins Terminal-Bench and OSWorld Verified (its 84.8 is self-reported). Kimi K3 was scored at max effort against xhigh for Muse Spark Open, so the gaps are not like for like. Choose Muse Spark Open for tool-call-heavy agents on a smaller deployment, and Kimi K3 for terminal and computer-use work.
Can you download Muse Spark Open and use it commercially?
Yes for most teams: the Muse Model License allows fine-tuning, distillation, merging and resale if you show "Built with Muse Spark Open" in each product's interface or documentation, and any model trained on its outputs counts as a Derivative with the same duty. Offering it as a Model-as-a-Service above US$20 million in aggregate gross revenue over any 12 months requires notifying Meta and signing a separate agreement. The Hugging Face download is gated, and the Usage Policy says the model is not intended for people under 18.
Does Muse Spark Open send your data to Meta?
A self-hosted copy runs wherever you deploy it, and the licence's only reporting duty is the Model-as-a-Service notice above. Meta claims no ownership of outputs. Hosted use is governed by each host's own retention and training terms, which differ by provider, so read those before sending sensitive prompts. Meta's card says its training data included material derived from Meta's own products and services.
Who should pick Muse Spark Open, and who should skip it?
Pick it for tool-using agents and screenshot-driven workflows where the weights must stay in your own account; Meta says one node of four Blackwell Ultra GPUs or eight H200s serves the NVFP4 build. Skip it if you need the lowest refusal rate, in which case GLM-5.3 or Kimi K3 fit better, or an Apache 2.0 licence, which Beam offers.
Top Alternatives
- Kimi K3: Pick Muse Spark Open for MCP-Atlas tool use on a smaller deployment; pick Kimi K3 for Terminal-Bench and OSWorld results in Meta's own table.
- GLM-5.3: Pick GLM-5.3 for the higher Intelligence Index and Terminal-Bench scores; pick Muse Spark Open for MCP-Atlas and native video and audio input.
- Beam: Pick Beam if you need a more permissive licence with no attribution rule; pick Muse Spark Open for its multimodal input and the wider set of day-one hosts.
- Muse Spark 1.3: Pick Muse Spark Open when weights must live on your hardware; pick Muse Spark 1.3 for Meta's own API tiers and a first-party price.