Muse Spark Open is free to download. Meta posted the weights on October 15, 2026 — the open-weight follow-up it had flagged as “coming soon” when it shipped closed-weight Muse Wonder three days earlier. On Hugging Face the checkpoint is meta-models/Muse-Spark-Open-816B-A42B: 816 billion total parameters, 42 billion active, a 1-million-token context window, and a natively multimodal stack (text, image, video and audio in; text out).
Hosted APIs are already live through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. On OpenRouter, checked the morning of October 16, the list price is $1.25 per million input tokens, $4.25 per million output and $0.15 for cached input — the same numbers Meta publishes for Muse Spark 1.3 Standard on its own Model API sheet. Meta’s research post does not publish a Meta-hosted token price for Spark Open; those figures are provider prices.
Meta’s research blog calls Muse Spark Open “the most capable US open-weight model to date.” That is a company claim. On Meta’s own open-weight comparison table, several Chinese open models still beat it on key rows. And to run the weights yourself, Meta’s model card sets a hard floor: at least 700 GB of GPU memory for the NVFP4 build, or 2 TB for BF16 — unless you pair two Apple M5 Ultra Mac Studios under MLX.
Muse Spark Open download: where the weights live
The primary download is Hugging Face, under the Muse Model License, in BF16 and NVFP4. Meta also lists Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Supported inference stacks, per the launch post and model card: SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth. Meta says the open release ships with custom Helion GPU kernels in SGLang that deliver 1.03–3.95× kernel-level speedups on NVIDIA and AMD GPUs, and a 3% end-to-end speedup on AMD MI350 — company claims, not third-party measurements.
Architecture, from the model card: a mixture-of-experts transformer with 256 experts and 8 active per token, plus a 1.8B Perception Encoder. Vocabulary size is 202,048. Context is 1M tokens at inference.
Hosted price vs the hardware bar
If you do not want to own the iron, third-party hosts are the path. OpenRouter’s list price for meta/muse-spark-open, checked October 16, 2026:
| Per 1M tokens (OpenRouter, Oct 16) | Muse Spark Open | Muse Spark 1.3 Standard (Meta’s sheet, for reference) |
|---|---|---|
| Input | $1.25 | $1.25 |
| Cached input | $0.15 | $0.15 |
| Output | $4.25 | $4.25 |
Those OpenRouter numbers match Meta’s published Muse Spark 1.3 Standard rates. Other hosts set their own prices. Self-hosting is a different bill: Meta’s model card says at least 700 GB of VRAM for NVFP4 or 2 TB for BF16. The desktop path Meta is pushing is two Apple M5 Ultra Mac Studios with MLX (up to 512 GB unified memory each). Meta says Spark Open is the first model in its class to run that way at home.
Meta’s open-weight table — and where Chinese models still win
Meta published its own comparison of Muse Spark Open against other open-weight models, all at xhigh reasoning effort for Spark Open. Every number below is from that Meta table. Do not treat them as independent, cross-lab results.
| Benchmark (Meta’s runs, xhigh) | Muse Spark Open | Qwen3.8-2.4T-A95B | GLM 5.3 | Kimi K3 |
|---|---|---|---|---|
| Artificial Analysis Intelligence Index | 39 | 40 | 45 | 44 |
| JobBench | 61.8 | 52.7 | 61.4 | 54.3 |
| MCP-Atlas | 89.2 | 84.5 | 84.2 | 82.3 |
| OSWorld Verified | 80.4 | — | — | 84.8 |
| Toolathlon Verified | 72.8 | 72.5 | 73.0 | 76.5 |
| HLE-Diamond (no tools) | 26.1 | 17.3 | 16.4 | 22.2 |
| GPQA Diamond | 89.0 | 93.5 | 91.7 | 93.5 |
| DeepSWE 1.1 | 60 | 57 | 67 | 68 |
| Terminal-Bench 2.1 | 76.4 | 82.0 | 83.9 | 85.0 |
Spark Open leads Meta’s table on JobBench, MCP-Atlas and HLE-Diamond (no tools) among the rows above. It does not lead the Artificial Analysis Intelligence Index, Terminal-Bench or DeepSWE — Qwen3.8, GLM 5.3 and Kimi K3 still sit ahead on those Meta-reported scores. That is the catch behind the “most capable US open-weight” line: even on Meta’s chart, the US lead is not a clean sweep.
License catch, Wonder contrast, and safety claims
The Muse Model License lets many developers use, modify, fine-tune, redistribute and commercialize the weights for free, with two hard conditions that matter in product work. You must prominently display “Built with Muse Spark Open.” And if you offer the model (or derivatives) as Model-as-a-Service and your aggregate gross revenue tops $20 million in any consecutive 12-month period, you must notify Meta and get a separate written agreement before continuing that MaaS offering. The usage policy also says the model is not intended for people under 18.
That is the opposite posture from Muse Wonder. Wonder launched October 12 as a closed-weight Model API product at $4/$20 per million tokens Standard, with Wonder Full available only through Muse Code plans and no public Full token price. Spark Open is the downloadable, multimodal MoE Meta said was coming. Meta also says Spark Open got additional safety post-training for the open release and met the thresholds in its Meta Superintelligence Scaling Framework; details are on the model card — company claims, not independent audits.
Next to watch: whether Meta publishes its own hosted token price for Spark Open, and whether independent labs reproduce the open-weight table.
Frequently asked questions
Where can I download Muse Spark Open?
Weights are on Hugging Face as Muse-Spark-Open-816B-A42B (BF16 and NVFP4), and Meta also lists Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Release date on the model card is October 15, 2026.
How much does Muse Spark Open cost?
The weights are free to download under the Muse Model License. Hosted APIs are priced by each provider. On OpenRouter, checked October 16, 2026, the list price is $1.25 per million input tokens, $4.25 output and $0.15 cached input. Meta has not published its own hosted token price for Spark Open on the research post.
What hardware do you need to run Muse Spark Open?
Meta’s model card says at least 700 GB of GPU VRAM for the NVFP4 build, or 2 TB for BF16. Meta also says it runs on two Apple M5 Ultra Mac Studios with MLX (up to 512 GB unified memory each).
What is the Muse Spark Open license catch?
Free use, modification, fine-tuning, redistribution and commercial use are allowed for many cases, but you must display “Built with Muse Spark Open.” If you offer it as Model-as-a-Service and your aggregate gross revenue exceeds $20 million in any consecutive 12 months, you must notify Meta and get a separate written agreement before offering it as MaaS. It is not intended for users under 18.
How is Muse Spark Open different from Muse Wonder?
Wonder (October 12) is closed-weight, sold on Meta’s Model API at $4/$20 Standard, with Full on Muse Code plans only. Spark Open (October 15) is the open-weight multimodal MoE Meta had called “coming soon,” free to download, with hosted access through third-party APIs.
What is Muse Spark Open’s context window?
1 million tokens at inference, per Meta’s model card. Vocabulary size is 202,048. Inputs are text, image, video and audio; output is text.
Sources
All accessed October 16, 2026, between about 8:40 and 8:50 AM ET.
- Meta research blog, “Introducing Muse Spark Open, The Most Capable US Open‑Weight Model to Date” (Oct 15, 2026; architecture, hosts, engines, Helion claims, safety, license summary, Apple MLX desktop claim): research.meta.ai/blog/muse-spark-open
- Hugging Face model card, meta-models/Muse-Spark-Open-816B-A42B (parameters, MoE layout, Perception Encoder, modalities, hardware bar, Meta open-weight benchmark table, Muse Model License text, safety notes; checked Oct 16, 2026): huggingface.co/meta-models/Muse-Spark-Open-816B-A42B
- OpenRouter models API / listing for meta/muse-spark-open (list price $1.25 / $4.25 / $0.15 cached; checked Oct 16, 2026): openrouter.ai/meta/muse-spark-open
- Meta pricing and rate limits (Muse Spark 1.3 Standard $1.25 / $0.15 / $4.25 for comparison; from prior coverage, pack 58): dev.meta.ai/docs/pricing-rate-limits
- Robotica Guide, Muse Wonder price / closed weights (Oct 13, 2026): roboticaguide.com/muse-wonder-price-api-closed-weights/
Related: Muse Wonder price, Full Code-only, weights closed · Mistral Large 4 price and open weights · More LLM news
Last updated: October 16, 2026. To report an error, see our corrections page. Articles are drafted with AI assistance and reviewed and edited by an editor; see our editorial policy.
