Muse Spark Open: Meta's 816B open-weight model, its licence terms, hardware needs and how it compares

Meta has released its largest open-weight Muse model so far. Muse Spark Open, published on 15 October, is a mixture-of-experts model with 816B total and 42B active parameters and a 1M-token context window, under Meta’s own Muse Model License. Meta calls it “the most capable US open-weight model to date”. Its demo shows the model running on two Apple Mac Studios.

This sits next to the closed Muse models we covered this week, Muse Wonder and Muse Spark on Meta’s Model API. Spark Open is a different product: you download the weights and run them yourself or through a hosting provider. Below: what the model is, what the licence actually asks of you, how it compares with the open models Indian teams already use, and what it takes to run.

The model in brief

From the Hugging Face model card:

• Size: 816B total parameters, 42B active per token. 256 experts, 8 routed per token, 62 layers.
• Inputs and outputs: text, image, video and audio in; text out. Vision goes through a 1.8B-parameter Perception Encoder.
• Context: up to 1M tokens at inference.
• Reasoning effort: low, medium, high or xhigh, with xhigh as the default.
• Weights: BF16 and NVFP4 checkpoints, plus community quantisations for llama.cpp, Ollama, LM Studio and Jan.
• Languages: trained on data from more than 100 languages. The card says it has not been evaluated in all of them and quality may drop outside the strongly supported set. It does not list which languages are strongly supported, so don’t assume Hindi, Tamil or Bengali quality without testing.

Meta’s launch post aims it at coding and agentic work: long-running tasks, software engineering and multi-step reasoning.

The licence: read section 3 if you sell API access

This is not an Apache or MIT release. The Muse Model License, as printed on the model card, allows use, fine-tuning, distillation, redistribution and sale, with conditions:

• Attribution. You must show “Built with Muse Spark Open” in your product’s interface, documentation, website or settings menu.
• Model-as-a-service threshold. If you offer model access to third parties through an API or similar, and you and your affiliates have more than US$20 million in gross revenue in any 12-month period, you must notify Meta and sign a separate agreement first. Internal use, research and using the model inside your own product are excluded.
• Usage policy. An attached policy bans categories such as malware, weapons, fraud and unlicensed financial, legal or medical practice.
• Outputs. Meta claims no ownership of what the model generates.

VentureBeat’s analysis also points out that “Derivatives” is defined broadly, including models improved using Spark Open’s outputs. If you plan to distil it into a smaller model, those smaller models carry the same obligations.

For most Indian startups and IT services firms, the US$20 million threshold won’t apply. It matters if you run an inference or GPU-cloud business, or if you belong to a large group whose combined revenue crosses it, because the test counts affiliates.

How it compares with other open models

Meta’s model card compares Spark Open with seven other open-weight models. The numbers below are Meta’s own, at the effort settings shown in its table, and have not been independently reproduced yet.

Benchmark (from Meta’s card)	Muse Spark Open	Qwen3.8 (2.4T-A95B)	GLM 5.3	Kimi K3	Mistral Large 4
Artificial Analysis Intelligence Index	39	40	45	44	38
MCP-Atlas (tool use)	89.2	84.5	84.2	82.3	–
JobBench	61.8	52.7	61.4	54.3	–
Terminal-Bench 2.1	76.4	82.0	83.9	85.0	–
DeepSWE 1.1	60	57	67	68	62
GPQA Diamond	89.0	93.5	91.7	93.5	–

Source: Muse Spark Open model card. A dash means Meta didn’t report a score.

What the table says:

• “Best US open model” is a narrower claim than “best open model”. On the overall index, the Chinese models GLM 5.3, Kimi K3 and Qwen3.8 score higher. VentureBeat lists more independent index scores for Chinese open models above 39.
• Its strengths are tool use and agent tasks. It leads the group on MCP-Atlas and JobBench.
• For terminal and repository coding, it trails. Kimi K3 and GLM 5.3 are ahead on Terminal-Bench 2.1 and DeepSWE 1.1.
• It refuses adversarial prompts far more often. Its StrongREJECT v2 score is 0.5 (lower is better); the other models in Meta’s comparison that report it scored between 62.9 and 86.4. The flip side, also on the card: its over-refusal score on FORTRESS (17.4) is higher than the others’, so expect it to decline more borderline but legitimate requests too.

For an Indian team, the practical question is often less about scores and more about origin. Some companies have procurement or client rules that prefer a US-origin model, and until now the US open options scored much lower: Meta’s table puts Nemotron 3 Ultra at 23 and Inkling at 26 on the same index. If that rule applies to you, Spark Open is a large step up.

Running it

Self-hosting. The card asks for at least 700GB of GPU memory for the NVFP4 weights and 2TB for BF16. Supported engines include SGLang, vLLM, TensorRT-LLM, Transformers, MLX and llama.cpp. Only 42B parameters are active per token, so it is fast for its size, but all 816B must sit in memory.

On Macs. Meta’s demo uses two M5 Ultra Mac Studios, each with up to 512GB of unified memory, linked to act as about a terabyte of memory under MLX (Meta). VentureBeat reports that the 512GB configuration ships later in October at an unannounced price. This is a striking demo, but it is not a cheap developer setup.

Hosted. For most teams, the realistic path is an API. Meta lists Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI, plus the weights in SageMaker, Google Cloud Model Garden and Microsoft Foundry. VentureBeat quotes Together’s serverless price as $1.25 per million input tokens and $4.25 per million output, and reports that context limits differ by provider (Groq at 131,072 tokens, Fireworks at the full 1M). Check each provider’s own page before you plan around a price or context size.

Data residency. None of the listed hosts says it serves the model from India. If your data must stay in India, self-hosting on Indian GPU capacity, or checking whether SageMaker or Google Cloud regions in Mumbai or Delhi can run it, is the route to explore.

Should you try it?

• Yes, if you build agents that call many tools, need long context with image or video input, or need a US-origin open model you can audit and fine-tune.
• Compare first, if your main workload is coding in a terminal or repository. On Meta’s own numbers, Kimi K3 and GLM 5.3 are ahead there.
• Test Indian languages yourself before relying on it for Hindi or regional-language users; the card makes no specific claims.
• Read the licence if you resell inference, and plan the “Built with Muse Spark Open” attribution into your UI.

The full evaluation report is still to come, Meta says. Independent results should follow over the next few weeks.