Meta ships Muse Spark Open, its answer to China’s best open-weight models
The New Stack Frederic Lardinois ● Covered by 18 sources
Meta just launched Muse Spark Open, an 816B open model with 42B active parameters. It’s aimed at China’s best open models, but the license still puts limits on big commercial users.
Based on reporting by The New Stack, Frederic Lardinois — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Meta launched Muse Spark Open on Thursday, and the company is calling it the most capable US open-weight model it has shipped so far. The model is huge on paper: 816B parameters, with 42B active at a time, plus a 1 million token window for long jobs and sprawling context.
The model is available through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. If someone wants the weights directly, Meta is also putting them on Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft’s Foundry Model Catalog. The pitch is open access, at least on the surface.
The license is where the picture gets messier. Meta is using its Muse Model License, which it says includes limited commercial terms. Companies with more than $20 million in gross revenue over any consecutive twelve-month period have to notify Meta and sign a separate written agreement before offering the model, or derivatives of it, as a service. Users also have to show “Built with Muse Spark Open” in the interface of products built on top of it.
Meta says the model is built for coding and agentic work, with strengths in long-horizon tasks, complex software engineering, and multi-step reasoning. In Meta’s own benchmark results, it beats Thinking Machines’ Inkling and Nvidia’s Nemotron 3 Ultra across the board, often by 20 points or more. Against Chinese open models, the fight is tighter: Spark Open leads Kimi K3, GLM 5.3, and Qwen3.8 on agentic and tool-use tests like JobBench and MCP-Atlas, and roughly ties them on SciCode.
Coding is where it falls behind. On Terminal-Bench 2.1, Kimi K3 scores 85.0 and GLM 5.3 scores 83.9, while Spark Open gets 76.4. On DeepSWE 1.1, the numbers are 68 for Kimi K3 and 60 for Spark Open. Meta had already seen a similar gap when it launched the closed version of Spark in April, and this open release doesn’t seem to have erased it.
Meta also says the model can run on two new Mac Studios with M5 Ultra chips and 512 GB of memory if they’re linked together with MLX. That setup is not exactly the sort of thing most people keep under the desk, but it did help Meta run Muse Code and ship updates to the Ice Cubes Mastodon client.
My take — AI-written commentary, not fact-checked reporting
This is the current open-model playbook in one neat package: generous weights, fancy benchmarks, and a license with strings attached for anyone who might actually make money. Open by download, semi-open by business terms — a very modern compromise. The real tell is the coding gap: Meta can beat Chinese rivals on some agent tasks and still not fully close the loop where developers live every day.
Read more about this at: The New Stack
Related stories
With new open models, Meta pitches another reboot of its struggling AI strategy
Ars Technica · 2 months ago ·
44
Meta Released Muse Spark 1.1 with Major Upgrades
The Neuron · 3 months ago ·
10
Meta says it has caught up with Anthropic and OpenAI with Muse Spark 1.3, its most powerful AI model yet
SiliconANGLE · 1 month ago ·
3