IMPLICATOR
Meta released the weights for Muse Spark Open on Thursday, a model it calls the most capable U.S. open-weight model to date. On Terminal-Bench 2.1, a coding test in Meta's own table, it scores 76.4 against Kimi K3's 85.0. It trails three Chinese open models, Kimi K3, GLM 5.3 and Qwen3.8, on that test.
The model card describes a mixture-of-experts model with 816 billion total parameters, 42 billion active per token, built for coding and agent work. It lists a million-token context window and text, image, video and audio inputs, with text output. Seven providers, including Together AI, host it; downloads span Hugging Face and Amazon, Google Cloud and Microsoft model catalogs.
Key Takeaways
- Meta released the weights for Muse Spark Open on Oct. 15, an 816-billion-parameter mixture-of-experts model with 42 billion active per token that it calls the most capable U.S. open-weight model.
- On Meta's own table it scores 76.4 on the Terminal-Bench 2.1 coding test against Kimi K3's 85.0, and 39 on the Artificial Analysis Intelligence Index, below GLM 5.3, Kimi K3 and Qwen3.8. None of the scores had been independently confirmed at launch.
- Apple's tests recorded 24.99 tokens per second on two M5 Ultra Mac Studios and 36.73 on four. The 512GB configuration is due in late October with no announced price.
- Companies whose gross revenue tops $20 million in any 12 months must sign a separate agreement with Meta before offering the model as a hosted service.
AI-generated summary, reviewed by an editor. More on our AI guidelines.
Chinese models lead on coding
GLM 5.3 scores 83.9 and Qwen3.8 82.0 on Terminal-Bench 2.1. On the Artificial Analysis Intelligence Index, Muse Spark Open's 39 trails GLM 5.3's 45, Kimi K3's 44 and Qwen3.8's 40, beating Inkling's 26 and Nemotron 3 Ultra's 23.
It leads the three Chinese models on JobBench, 61.8 against GLM 5.3's 61.4, and MCP-Atlas tool use, 89.2 against Qwen3.8's 84.5.
Meta produced and published all these scores, including the Intelligence Index figures, in its model card table; none had been independently confirmed for the released model at launch.
“These aren't the most capable models that we've developed or released,” Meta chief AI officer Alexandr Wang said. “But we think that there will be a lot of companies and developers who sort of use this model and tinker with it at small scale and then seek to ramp up and scale up their usage.”
FREE · ABOUT FIVE MINUTES
Follow the open-weight model race as it moves.
Implicator cuts the flood to the developments worth your attention, with concise reporting on what changed and why it matters. Delivered every weekday at 4:45 a.m. Pacific, 7:45 a.m. Eastern.
From San Francisco. No spam. Unsubscribe anytime.
Muse Spark Open is less advanced than Muse Wonder. Muse Wonder and Muse Wonder Full, released Monday, remain closed.
From May's safety concerns to release
In an interview released May 13, Wang said Muse Spark had triggered safety checks that made it “not suitable for open sourcing”. In August, alongside the 30-billion-parameter Muse Glimmer, Meta promised to publish Muse Spark 1.2 weights in the coming weeks. This release came about two months later.
The open version received additional safety post-training for use without system guardrails. Meta says it met the release thresholds of its Meta Superintelligence Scaling Framework. That claim rests on Meta's own evaluation; it had not published the full report at launch.
Running it on Macs
Meta worked with Apple to run the model through MLX on two M5 Ultra Mac Studios with 512GB of unified memory each, linked over Thunderbolt 5, or four with 256GB each.
Apple's October 15 tests recorded 24.99 tokens per second on two machines and 36.73 on four, both using tensor parallelization, 4-bit NVFP4 weights, an 8,192-token context and 128 generated tokens per test. Apple says four Mac Studios generate over 36 tokens per second, “well above average human reading speed.” Those short generation tests do not establish how quickly the model completes a sustained coding task or performs at its full context window.
Know someone who'd find this useful? ✉️ Email it to a friend in one click, or they can subscribe free here.
The 512GB configuration is due in late October and has no announced price. The 256GB version costs at least $10,799, putting four above $43,000.
License terms and hosted prices
Meta describes the Muse Model License as allowing many developers free use. Companies whose gross revenue, including affiliates, exceeds $20 million in any consecutive 12 months must notify Meta and sign a separate agreement before offering it as a hosted service. Internal use and incorporation into their own products are exempt.
Products must display “Built with Muse Spark Open.” Military and critical-infrastructure uses are barred.
Together AI lists launch pricing at $1.25 per million input tokens and $4.25 per million output tokens, matching Meta's hosted Spark rates since July. Z.AI lists GLM 5.3 at $1.40 per million input tokens and $4.40 per million output tokens.
“I waited so long for US open weight models to be relevant again,” Interconnects author Nathan Lambert wrote on X. He added: “Still more work to do!”
Frequently Asked Questions
What is Muse Spark Open?
An open-weight model Meta released on Oct. 15, 2026. It uses a mixture-of-experts design with 816 billion total parameters, 42 billion of them active per token, and a million-token context window. It accepts text, image, video and audio input, produces text, and was built for coding and agent work.
How does it compare with Chinese open models?
In Meta's own table it scores 76.4 on the Terminal-Bench 2.1 coding test, behind Kimi K3 at 85.0, GLM 5.3 at 83.9 and Qwen3.8 at 82.0. It leads those three on JobBench and MCP-Atlas tool use. Meta produced all the scores, and none had been independently confirmed at launch.
Can it run on a Mac?
Meta worked with Apple to run it through MLX on two M5 Ultra Mac Studios with 512GB of unified memory each, or four with 256GB each. Apple's tests recorded 24.99 tokens per second on two machines and 36.73 on four. The 512GB configuration is due in late October, and the 256GB version costs at least $10,799.
Is it free for commercial use?
Meta says the Muse Model License allows many developers free use. Companies whose gross revenue, including affiliates, exceeds $20 million in any consecutive 12 months must notify Meta and sign a separate agreement before offering it as a hosted service. Products must display "Built with Muse Spark Open," and military and critical-infrastructure uses are barred.
Why did Meta release Muse Spark weights after saying it would not?
In May, Alexandr Wang said Muse Spark had triggered safety checks that made it "not suitable for open sourcing." Meta says the open version received additional safety post-training and met the release thresholds of its Meta Superintelligence Scaling Framework. It had not published the full evaluation report at launch.
AI-generated summary, reviewed by an editor. More on our AI guidelines.
Free AI briefing · Weekdays
The AI stories that matter, sourced and explained.
Join the Morning Briefing. It goes out every weekday at 4:45 a.m. Pacific, with later sends for the East Coast, Berlin and Tokyo.
Free when you sign up: The Paperclip Compendium, our tested guide to running AI agents.
Free. Unsubscribe in one click.
Marcus Schuler
Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai
