Meta Muse Spark Open: 816B Open Weights, License, Specs

October 16, 2026 | AI | By Mahesh

Meta Platforms (NASDAQ: META) released Muse Spark Open on October 15, 2026, an 816 billion parameter open-weight model that Meta calls "the most capable US open-weight model to date" [1]. It uses 42 billion active parameters per token, has a 1 million token context window and ships under a custom Muse Model License rather than a standard open-source license [1][2]. Meta says the model runs on two Apple M5 Ultra Mac Studios. Its own benchmark table also shows it behind the Chinese open models GLM 5.3 and Kimi K3 on several tests.

What is Muse Spark Open and what is new

Muse Spark Open is an open-weight mixture-of-experts model from Meta Superintelligence Labs, built for coding and long-running agent tasks and released with downloadable weights on October 15, 2026 [1][2]. The weights are on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Hosted APIs come from Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI [1].

Meta demonstrated the model running locally through Muse Code, shipping iOS changes to the open-source Mastodon client Ice Cubes and driving the video editor DaVinci Resolve to find and crop strong moments from footage [1]. It follows Meta's closed Muse Wonder, which costs $4 and $20 per million tokens through Meta's API and has no downloadable weights, as covered in Meta Muse Wonder Launch: 1M Context, $4 API Price, Benchmarks. Bloomberg framed the pair as Meta releasing an open model while keeping its best technology closed [7].

"I waited so long for US open weight models to be relevant again"

Nathan Lambert, author at Interconnects, on X, quoted by VentureBeat [4]

Lambert followed that line with "Still more work to do," and the benchmark section shows why [4].

Under the hood: architecture and hardware

Muse Spark Open has 816 billion total parameters and 42 billion active per token, so about 5% of the weights fire on each token. It is a mixture-of-experts transformer with 62 layers, a hidden size of 8,192 and 256 experts, 8 of which are activated per token. Attention uses 128 query heads and 16 key-value heads, and the vocabulary holds 202,048 tokens [2].

A 1.8 billion parameter Perception Encoder handles vision. The model accepts text, image, video and audio input and produces text only. It was trained on data from more than 100 languages, though Meta says not all were evaluated. Reasoning effort has four settings, low, medium, high and xhigh, with xhigh as the default and the recommended setting for coding and agent work [2].

The model card lists the hardware floor at 700 GB of accelerator memory for the NVFP4 build and 2 TB for BF16. Meta's desktop demo uses two M5 Ultra Mac Studios with 512 GB of unified memory each through the MLX framework [1][2]. VentureBeat notes the 512 GB configuration ships in late October at an unannounced price, and that a 256 GB M5 Ultra lists at $10,799, so four of them would cost $43,196 before tax [4]. The Hub lists 12 quantized variants for llama.cpp, Ollama, LM Studio and Jan.

Benchmarks: Meta's table against outside listings

The scores below are Meta-reported at xhigh reasoning effort, and its evaluation methodology report [3] holds the settings. The coding results use each model's preferred harness, so the agent software is part of the comparison [4].

Muse Spark Open leads on JobBench and on MCP-Atlas, a test of tool use. It trails on the coding and economic-task rows, scoring 76.4 on Terminal-Bench 2.1 against 85.0 for Kimi K3 [2][4]. The model card also reports wins on HLE-Diamond without tools (26.1 against 22.2 for Kimi K3) and on ZeroBench with tools (50.0 against 41.0) [2].

The headline index is the weak point. VentureBeat says Artificial Analysis could not verify Meta's score of 39 at publication, and that its independent listings put MiMo-V2.6-Pro at 46, GLM-5.3 at 45, Kimi K3 at 44, GLM-5.3-Flash at 42 and Qwen3.8 at 40, all above Meta's figure [4]. Meta also reports safety scores, including a 99.2% jailbreak refusal rate on its internal cybersecurity benchmark against 75.5% for Kimi K3. The model card warns those results apply only to the released checkpoint and may not hold for modified models, which matters for an open release [2].

Pricing, availability and license terms

The weights cost nothing to download, but access is gated. Hugging Face asks for a full legal name, date of birth and organization details before it releases them [2]. Meta has not published hosted token rates. Together AI lists $1.25 per million input tokens, $4.25 per million output tokens and $0.15 for cached input, against $1.40 and $4.40 for GLM-5.3, $2 and $6 for Qwen3.8-Max and $3 and $15 for Kimi K3 [4].

The Muse Model License permits use, fine-tuning, distillation, merging and redistribution, and Meta claims no ownership of outputs [2][5]. Products using the model must display "Built with Muse Spark Open." A company that earns more than $20 million in gross revenue in any 12 months from offering model access to third parties must notify Meta and sign a separate agreement. Internal use, research and embedding the model in your own product are exempt. The acceptable use policy bars weapons and military use, critical infrastructure operation, malware and use by anyone under 18 [2][5].

Derivatives are defined broadly and include models improved using Spark Open outputs, and derivatives carry the same attribution rule [4][5]. This is a custom license, not Apache 2.0 or MIT. The official weights, model card and discussion board are at the Hugging Face model page.

Limitations, trade-offs and alternatives

Running it yourself is expensive. The 700 GB NVFP4 floor means a multi-GPU server for most teams, and Meta's two-Mac setup has no published price or measured token speed [2][4]. Hosts also trim the context window. Groq lists 131,072 tokens while Fireworks lists 1,048,576, so the 1 million token figure is not guaranteed everywhere [4]. Output is text only, and the card notes reasoning errors in novel multi-step tasks and weaker support for some languages [2].

Meta's table compares the model with Inkling, Nemotron 3 Ultra, Beam, Mistral Large 4, Qwen3.8, GLM 5.3 and Kimi K3 [2]. Teams that want weights have several routes. GLM 5.3 and Kimi K3 score higher on the independent index, Mistral's open model is detailed in Mistral Large 4 Preview: 1.05T Open-Weight MoE, Specs and Pricing, and Nvidia released Nemotron 3.5 Super VL, a 120 billion parameter open model with 12 billion active parameters, on October 15 [6]. Beam, another model in Meta's comparison set, comes from Nvidia-backed Reflection AI, and Depth Grid covered the Nvidia talks in Nvidia in Talks to Deepen Stake in or Buy Reflection AI.

Sources

1. Meta Superintelligence Labs. "Introducing Muse Spark Open, The Most Capable US Open-Weight Model to Date," October 15, 2026. research.meta.ai
2. Meta. "Muse-Spark-Open-816B-A42B," model card, Hugging Face. huggingface.co
3. Meta. "Muse Spark Open evaluation methodology." research.meta.ai
4. VentureBeat. "Meta unveils Muse Spark Open, the new top U.S. open weights model, and it runs on two paired Mac desktops." venturebeat.com
5. Trending Topics. "Meta's Muse Spark Open Comes With Strings Attached." trendingtopics.eu
6. AI Weekly. "AI News Today, October 15: Top Stories." aiweekly.co
7. Bloomberg. "Meta Releases Open AI Model While Keeping Its Best Tech Closed," October 15, 2026. bloomberg.com

Read More on Depth Grid

• Meta Muse Wonder Launch: 1M Context, $4 API Price, Benchmarks
• Mistral Large 4 Preview: 1.05T Open-Weight MoE, Specs and Pricing
• Claude Haiku 5.5 Launch: $0.10 Input, 1M Context, Benchmarks
• Tech News Today, Oct 14: Character.AI, OpenAI, Meta, Bumble

Article by Mahesh | Depth Grid