Muse Spark Open is Meta’s new open-weight AI model, released October 15, 2026, with 816 billion total parameters, 42 billion active per token and a 1-million-token context window. Weights are free on Hugging Face; hosted access lists at $1.25 per million input tokens. By Meta’s own benchmarks, it still trails China’s Kimi K3 and GLM 5.3 on coding.
Key takeaways
- Muse Spark Open has 816B parameters, but only 42B fire on each token.
- Kimi K3 scores 85.0 on Terminal-Bench 2.1; Muse Spark Open scores 76.4.
- API resellers with over $20 million in revenue need a separate deal with Meta.
What is Muse Spark Open?
Muse Spark Open is the downloadable version of Meta’s Muse Spark model, built for coding and AI agents. Meta published the weights on October 15 under its own Muse Model License. It is a mixture-of-experts model: 816 billion parameters in total, with 42 billion active at any one time.
Meta’s research team calls it “the most capable US open-weight model to date” in its launch post. That is a careful claim. It says US, not global.
According to the PyTorch team at Meta, the open model “shares the same base model as Muse Spark 1.3,” the closed version Meta sells through its own products. The open release also got “additional safety post-training to make it more robust against misuse as a standalone model without system guardrails.”
The specs that matter
- Architecture: 62 layers and 256 experts, with 8 active per token, according to Unite.AI’s read of the model card.
- Inputs: text, image, video and audio in; text out.
- Context: up to 1 million tokens.
- Reasoning effort: low, medium, high and xhigh, with xhigh as the default.
- Languages: trained on more than 100.
This is Meta’s second open-weight Muse release. Reuters noted that Meta shipped the smaller Muse Glimmer in August.
How good is Muse Spark Open against Kimi K3 and GLM 5.3?
It is strong on tool use and weaker on raw coding. On Meta’s own model card, Muse Spark Open leads MCP-Atlas with 89.2 but trails Kimi K3 and GLM 5.3 on Terminal-Bench 2.1 and on the Artificial Analysis Intelligence Index. The best open models in Meta’s own comparison are still Chinese.
Every number below is self-reported by Meta or taken from its model card. None has been independently reproduced yet. Meta says a full evaluation report will be available “soon.”
| Benchmark | Muse Spark Open | Kimi K3 | GLM 5.3 | Qwen3.8 |
|---|---|---|---|---|
| Artificial Analysis Intelligence Index | 39 | 44 | 45 | 40 |
| Terminal-Bench 2.1 (coding agents) | 76.4 | 85.0 | 83.9 | — |
| MCP-Atlas (tool use) | 89.2 | 82.3 | 84.2 | — |
| GPQA Diamond (science) | 89.0 | 93.5 | — | 93.5 |
| DeepSWE 1.1 | 60 | 68 | — | — |
Where Meta does lead
Safety is Meta’s pitch. The model card lists a 99.2% jailbreak refusal rate on Meta’s internal cyber benchmark, against 75.5% for Kimi K3 in the same setup. It also reports 98.1% honesty on the MASK test.
Meta also leads its closest US rival by a wide margin. OrcaRouter reported that Nvidia’s Nemotron 3.5 Super VL scores 26.49 on the same intelligence index, against 39 for Muse Spark Open. If you need a US-built open model, Meta now sits on top.
How much does Muse Spark Open cost to run?
The weights are free to download. Hosted access through OpenRouter lists at $1.25 per million input tokens and $4.25 per million output tokens, with cached input at $0.15, according to a capture by Digital Applied on October 16. Running it yourself means buying or renting hundreds of gigabytes of memory.
That $1.25/$4.25 price matches what Meta already charges for its hosted Muse Spark 1.2, per OrcaRouter. In other words, the open model costs the same through an API as the closed one. The savings only appear when you self-host at scale.
Artificial Analysis put the cost at $0.84 per completed task on its Intelligence Index run, OrcaRouter reported. The model used about 110 million tokens on that run, below the 130 million median.
The hardware bill
| Build | File size | What you need |
|---|---|---|
| BF16 (full precision) | ~1.6 TB | About 2 TB of GPU memory |
| NVFP4 (quantized experts) | ~530 GB | About 700 GB of GPU memory, e.g. 8 H200s |
| NVFP4 on Apple MLX | Fits in 1 TB | Two M5 Ultra Mac Studios with 512 GB each |
| Unsloth GGUF (~2-bit) | ~300 GB | Combined RAM and VRAM |
The two-Mac demo is the headline Meta wants. Each M5 Ultra Mac Studio offers up to 512 GB of unified memory, Meta said. But the Mac route runs a quantized build, not the full model. The full-precision model needs a data-center rack.
What does the Muse Model License allow?
It is open weights, not open source. You can fine-tune, distill, merge and sell derivatives, and Meta claims no ownership of outputs. But every product must display “Built with Muse Spark Open,” and API resellers above $20 million in annual revenue must sign a separate agreement with Meta first.
According to OrcaRouter’s review of the license, the Hugging Face repository is gated. The request form asks for your full legal name, date of birth and organization, and it cannot be edited after you submit it.
The $20 million clause is aimed squarely at cloud resellers. Internal use, research and building the model into your own product are exempt, Digital Applied reported. A startup shipping a coding tool can use it freely. A hosting company selling raw tokens at scale cannot, at least not without calling Meta.
Compare that with Reflection AI’s Beam, which shipped under Apache 2.0. We covered that launch in Reflection Beam Ships 501B Open Weights, 23B Active. Apache 2.0 has no attribution banner and no revenue trigger.
Who wins and who loses from Muse Spark Open?
Inference hosts win first. Seven API providers launched with the model on day one, and three clouds list the weights. Chinese labs lose some of their claim to “best open model you can legally use in the US.” Closed-model vendors face another price anchor near the bottom of the market.
Winners
- Inference providers: Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI all serve it. Every hosted token is their revenue, not Meta’s.
- Hyperscalers: Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry list the weights. A 1.6 TB model sells a lot of GPU hours.
- Apple: Meta chose two M5 Ultra Mac Studios as its desktop showcase. That is free marketing for Apple’s high-memory machines.
- US enterprises wary of Chinese models: they now have a domestic option at 39 on the index, up from Nemotron’s 26.49.
Losers and open questions
Meta itself does not earn directly from the weights. It expects capital spending of $130 billion to $145 billion this year, according to TIKR’s summary of its guidance. Giving away a model trained on that budget only pays off if Muse becomes the default developers build on.
That logic worked for Meta’s ads business, which is the cash engine here. Meta’s second-quarter revenue rose 28% to $60.8 billion, TIKR reported. But free cash flow in that quarter was just $784 million. Investors will want to know how long that gap lasts.
Here is the skeptical read. The open model is not Meta’s best. OrcaRouter puts the closed Muse Spark 1.3 at 48 on the same index, against 39 for the open version. Meta is giving away its second-best model and keeping the top one behind a paywall.
That fits Meta’s pattern. Its closed Muse products are where it charges: we covered the $20-a-month Meta Muse AI agent in September and Muse Wonder’s pricing on October 12.
Should developers switch to Muse Spark Open?
Only for specific jobs. If you need agentic tool calling, long context and a US-origin model you can host privately, it is now the strongest option. If you want the best open coder regardless of origin, Kimi K3 and GLM 5.3 still score higher on Meta’s own chart.
Early partners point to tool use. “Muse Spark Open consistently selected the right next tool and kept its responses compact,” said Zimu Li of Factory, quoted on Meta’s developer blog. Ollama’s Dong Chen said it “performs particularly well on tool calling and long context.”
Those are launch partners, so weigh them accordingly. For a deeper cost comparison with closed models, see our earlier test of Muse Spark vs Claude Opus 5.
What to watch next
The next hard date is Meta’s third-quarter earnings on October 21. After that, the signals are the full evaluation report, independent benchmark runs and whether hosted prices fall below Meta’s own $1.25 list.
- October 21, 2026: Meta reports Q3 results after the market close, with the call at 4:30 p.m. ET, per Meta Investor Relations. Listen for capex guidance and any Muse adoption numbers.
- Full evaluation report: Meta promised it “soon” with no date. Independent scores could confirm or shrink the 89.2 MCP-Atlas lead.
- Hosted price war: watch whether Groq, Together AI or Fireworks undercut $1.25 per million input tokens.
- Groq general availability: Groq launched in limited preview. Wider access would test Meta’s speed claims.
FAQ
Is Muse Spark Open free?
The weights are free to download from Hugging Face after a gated sign-up. Hosted access costs $1.25 per million input tokens and $4.25 per million output tokens on OpenRouter.
Is Muse Spark Open open source?
No. It is open weights under the Muse Model License, which requires a “Built with Muse Spark Open” notice and a separate deal for API resellers above $20 million in revenue.
Can I run Muse Spark Open on a Mac?
Yes, with two M5 Ultra Mac Studios carrying 512 GB of unified memory each, running a quantized NVFP4 build through Apple’s MLX framework.
How many parameters does Muse Spark Open have?
It has 816 billion total parameters, with 42 billion active per token, plus a 1-million-token context window.
Is Muse Spark Open better than Kimi K3?
Not on coding. Kimi K3 scores 85.0 on Terminal-Bench 2.1 versus 76.4 for Muse Spark Open. Meta’s model leads on MCP-Atlas tool use, 89.2 to 82.3.
Where can I use Muse Spark Open?
Through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI, or via Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry.
The bottom line
Muse Spark Open is the best US open-weight model on paper, and a real option for agent builders who cannot use Chinese models. It is not the best open model in the world, by Meta’s own numbers, and it is not Meta’s best model either. For investors, the release is a distribution play paid for by a capex budget of up to $145 billion. The October 21 earnings call is where Meta has to show it is working.
Free guide · 12 pages
AI Tools Every Investor Should Use in 2026
10 tools for research, stock signals, charts and crypto, what each one costs, three ready-made stacks from $0 a month and 7 copy-paste prompts for filings and earnings calls. Enter your email and we will send it to you.
Plus one email a week: AI Money This Week, every Sunday. Unsubscribe anytime.
Sources
- Meta AI Research: Muse Spark Open
- PyTorch: Muse Spark Is Now an Open-Weights Model
- Meta AI: Build with Muse Spark Open
- Reuters via Investing.com: Meta releases open-weight Muse Spark model
- Unite.AI: Meta’s 816B-Parameter Muse Spark Open Weights Are Now Public
- Digital Applied: Muse Spark Open licence and hardware
- OrcaRouter: Muse Spark Open and the licence fine print
- TIKR: Meta stock and the open-weight Muse Spark model
- Meta Investor Relations: Q3 2026 results date
Wealth Engine researches and drafts with AI tools and checks every figure against the sources above. How we report.
Leave a Reply