Meta releases "America's strongest open-source model"

Author: Su Yang, Tencent Technology

On October 15th, US local time, Meta officially released Muse Spark Open, a hybrid expert model with 816 billion total parameters and 42 billion active parameters, and also made the weights available for download.

This is the most powerful open-source weight model in the United States to date, natively supporting multimodal inputs and with a context window of up to 1 million tokens.

Turing Award winner and former Meta Chief AI Scientist Yang Likun gave it a "passionate thumbs up".

Two "Apple" machines can run

Meta claims to have collaborated with Apple to optimize the system, allowing two M5 Ultra Mac Studio instances to run the entire model.

The specific configuration consists of two M5 Ultra Mac Studio instances, working collaboratively through the MLX framework, with a unified memory capacity of 1TB. The M5 Ultra's memory bandwidth reaches 1.2TB per second, which can support the small slice read speed required for each token.

The model itself employs a hybrid expert architecture, with 42 billion parameters activated in each iteration out of a total of 816 billion. This means that the actual number of parameters invoked during inference is far less than the total size, but all the full weights still need to be loaded into memory. The NVFP4 version, as indicated on the model card, requires at least 700GB of accelerator memory, while the BF16 version requires 2TB.

Therefore, two 512GB Mac Studios just barely met the threshold of being able to store the heavy quantitative copyright content.

Native multimodality is another key feature of Muse Spark Open.

The model accepts four types of input: text, images, video, and audio, and outputs text. It is not a generative multimodal model—it cannot draw, generate videos, or synthesize speech; its strength lies in understanding and analyzing information from multiple modalities. Developers can use it to view screenshots, listen to recordings, analyze video frames, and then combine this with text commands to make judgments. For proxy-based workflows, this understanding capability is more practical than its generative capability.

In addition to Apple chips, Muse Spark Open also supports mainstream inference engines and hardware platforms. SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, MLX, llama.cpp, and Unsloth are all compatible, and chips from AMD, Apple, Intel, and Nvidia can run.

Meta also comes with a custom Helion GPU kernel, which can deliver 1.03 to 3.95 times the kernel-level speedup on SGLang and 3 times the end-to-end speedup on AMD MI350.

The true value of "America's strongest"

Meta has repeatedly emphasized that it is "the strongest open-source weight model in the United States".

In the benchmark tables published by Meta itself, Muse Spark Open leads in the JobBench and MCP-Atlas tool usage benchmark tests, scoring 61.8 and 89.2 points respectively, higher than competitors such as GLM 5.3, Kimi K3 and Qwen3.8.

In the HLE-Diamond reasoning test without tools, it also led with a score of 26.1. In the visual evaluation with tools, ZeroBench scored 50 points and SimpleVQA scored 73.6 points, also ranking highly.

American AI scientist Nathan Lambert said he has been waiting a long time for American open-source models to become competitive again, and now he finally sees hope.

Box CEO Aaron Levie also believes this is a major step forward for the US open-source community, allowing companies to mix and match – using cutting-edge models for orchestration and high-level tasks, and low-cost custom models for repetitive workflows.

There is a lot of support, but the shortcomings are also obvious.

In the Terminal-Bench 2.1 terminal encoding test, Muse Spark Open scored 76.4 points, lower than GLM 5.3's 83.9 points and Kimi K3's 85.0 points.

On the software engineering task DeepSWE 1.1, a score of 60 was also behind GLM and Kimi's scores of 67 and 68.

On the Comprehensive Knowledge Work Assessment (GDPVal-AA) v2, Muse Spark Open scored 1401 points, which is lower than GLM 5.3's 1653 points, Qwen 3.8's 1613 points, and Kimi K3's 1537 points.

The intelligence index from the third-party testing organization Artificial Analysis also provides a reference.

According to its publicly available rankings, Muse Spark Open (Xhigh mode) scored 39, Xiaomi MiMo-V2.6-Pro scored approximately 46, Zhipu GLM-5.3 scored approximately 45, Kimi K3 scored approximately 44, GLM-5.3-Flash scored approximately 42, and Alibaba Qwen3.8 2.4T A95B scored approximately 40.

In other words, at least six released open-source models have a higher overall score than Muse Spark Open, and all of them are from Chinese vendors.

Over the past year, downloads of Chinese open-source models have surpassed those of US models. On OpenRouter, 60% of tokens from US origin requests are processed by open-source models, with the majority originating from Chinese labs. This release by Meta sets a new benchmark within the US open-source community.

In terms of pricing, Together AI announced that the Muse Spark Open API is priced at $1.25 per million input tokens and $4.25 per million output tokens, which is the same as the closed-source version of Muse Spark.

This price isn't cheap for an open-source model. DeepSeek-V4.1-Flash costs only $0.75 per million tokens during off-peak hours, Xiaomi MiMo-V2.6-Pro costs $1.305, and GLM-5.3 costs $5.80.

Muse Spark Open is positioned between the mid-range and high-end markets, and its cost-effectiveness advantage is mainly reflected in self-hosted scenarios—if enterprises deploy it themselves, the cost of computing power may be more cost-effective than API calls after being spread out.

From fixing iOS devices to editing videos

The parameters and benchmark scores are just data on paper; what truly gives you a sense of significance are the several real-world use cases of Muse Spark Open.

Meta's first published use case is modifying an iOS app.

Meta used Muse Code (a programming agent based on Muse Spark Open) to completely process Ice Cubes—an open-source Mastodon client—locally. Specific tasks included bug fixing, adding new features, and completing a full build.

The second use case is Muse Spark Open driving DaVinci Resolve locally for video editing, reviewing footage frame by frame, finding the most exciting segments, and then cropping them into a vertical format.

Traditional video editing tools mostly perform basic tasks such as automatically adding subtitles and simple transitions, while Muse Spark Open makes more subjective decisions such as "determining which shots are best." It can view each frame, understand the content of the shots, make selections, and then directly operate professional editing software to output the final product.

If this process is successfully implemented, it will directly improve efficiency for short video production, content creators, and media organizations.

Muse Spark Open also supports agent shopping. Pedro Cuenca, a machine learning engineer at Hugging Face, shared a shopping demo on X, showcasing Spark Open's capabilities in agent tasks, computer usage, and multimodal reasoning.

Specifically, this means the model can directly manipulate the browser, browse e-commerce websites, compare products, and complete the purchase. This scenario of "AI helping you buy things" has been discussed for a long time, but it has always been limited by the model's ability to understand web pages and its operational stability. Spark Open's multimodal understanding and tool-calling capabilities have significantly lowered the barrier to entry for such applications.

Sophia Yang, the head of developer relations at Fireworks AI, demonstrated an even cooler approach—using Spark Open to generate code frame by frame to create a complete interactive animated webpage.

Users can browse the entire webpage as if stepping into a movie scene, with each frame written in real-time by the model. This demo is more of a technical showcase, but it demonstrates the possibilities of combining large local models with creative tools.

Behind these use cases lies Meta's complete layout on the open-source model.

In April of this year, the first-generation Muse Spark was released, followed by Glimmer with 30 billion parameters and the Apache 2.0 license in August, and the more advanced Muse Wonder and Wonder Full in early October. Now Spark Open has filled the gap in open-source weight, forming a complete product line from small local models to large closed-source models.

The licensing agreements are also segmented. Glimmer uses the completely open Apache 2.0, while Spark Open uses the custom Muse Model License—most developers can use it for free, but if your company provides Model as a Service and has annual revenue exceeding $20 million, you will need to sign a separate agreement with Meta.

Internal deployments, research, and commercial applications that are not model services are not within the scope of the restrictions. This design is very precise—it's fine for internal use within an enterprise, and it's also fine for using it to develop your own products, but if you want to sell the API directly, then you have to negotiate a revenue share.

Source:Global Cybersecurity Alliance (GCSA)
Website:www.gcsa.org