Meta Releases 'America's Strongest Open Source Model'

Meta CEO Mark Zuckerberg. Image processed by AI

Written by: Su Yang

Edited by: Xu Qingyang

On October 15, local time in the United States, Meta officially launched Muse Spark Open, a mixed expert model with a total of 816 billion parameters and 42 billion active parameters, while also making the weights available for download.

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

Turing Award winner and former Chief AI Scientist at Meta, Yang Likun, enthusiastically praised the model.

01 Two "Apples" Can Run It

Meta stated that it collaborated with Apple to optimize the model, which can be fully run on two M5 Ultra Mac Studios.

The specific configuration involves two M5 Ultra Mac Studios working together through the MLX framework, with a unified memory of 1TB. The memory bandwidth of the M5 Ultra reaches 1.2TB per second, supporting the small slice read speed required for each token.

The model itself adopts a mixed expert architecture, activating 42 billion parameters at a time out of the total 816 billion. This means that the actual number of parameters called during inference is much smaller than the total scale, but the complete weights still need to be loaded into memory. The NVFP4 version marked on the model card requires at least 700GB of accelerator memory, while the BF16 version requires 2TB.

Thus, two 512GB Mac Studios are just at the threshold of being able to accommodate the quantized weights.

Native multimodality is another key feature of Muse Spark Open.

The model accepts four types of input: text, images, videos, and audio, with the output being text. It is not a generative multimodal model—it cannot draw, generate videos, or synthesize speech; its capability lies in understanding and analyzing information from various modalities. Developers can make it view screenshots, listen to recordings, analyze video frames, and then make judgments based on textual instructions. For agent-based workflows, this understanding capability is more practical than 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 have all been adapted, and chips from AMD, Apple, Intel, and NVIDIA can run it.

Meta also includes a custom Helion GPU kernel, which can provide 1.03 to 3.95 times kernel-level acceleration on SGLang, and a 3% end-to-end speedup on AMD MI350.

02 The Value of "America's Strongest"

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

In the benchmark scores released by Meta, Muse Spark Open leads in two tool-based benchmark tests, scoring 61.8 and 89.2 in JobBench and MCP-Atlas, respectively, surpassing competitors like GLM 5.3, Kimi K3, and Qwen3.8.

In the HLE-Diamond inference test without tools, it also leads with a score of 26.1. In tool-assisted visual assessments, ZeroBench scored 50, and SimpleVQA scored 73.6, also ranking at the top.

American AI scientist Nathan Lambert stated that he has been waiting for American open-source models to regain competitiveness for a long time, and he finally sees hope now.

Box CEO Aaron Levie also believes this is a significant advancement for the American open-source community, allowing companies to mix and match—using cutting-edge models for orchestration and high-level tasks while employing low-cost custom models for repetitive workflows.

The enthusiasm is high, but the shortcomings are also evident.

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

In the software engineering task DeepSWE 1.1, it scored 60, also trailing behind GLM and Kimi's scores of 67 and 68.

In the comprehensive knowledge work assessment GDPVal-AA v2, Muse Spark Open scored 1401, which is lower than GLM 5.3's 1653, Qwen3.8's 1613, and Kimi K3's 1537.

The intelligent index provided by the third-party testing agency Artificial Analysis also offers a reference.

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

This means that at least six released open-source models have comprehensive scores above Muse Spark Open, all from Chinese manufacturers.

In the past year, the download volume of Chinese open-source models has surpassed that of American models, with 60% of the token volume in requests from the United States on OpenRouter being processed by open-source models, most of which come from Chinese laboratories. Meta's release this time sets a new benchmark in the American open-source camp.

In terms of pricing, Together AI announced that the pricing for the Muse Spark Open API is $1.25 per million input tokens and $4.25 for output, which is on par with the closed-source version of Muse Spark.

This price is not considered cheap among open-source models; DeepSeek-V4.1-Flash costs only $0.75 per million tokens during non-peak hours, Xiaomi MiMo-V2.6-Pro is $1.305, and GLM-5.3 is $5.80.

Muse Spark Open is positioned between mid-range and high-end, with its cost-performance advantage mainly reflected in self-hosted scenarios—if enterprises deploy it themselves, the computational cost may be more economical than API calls.

03 From Fixing iOS to Video Editing

Parameters and benchmark scores are just paper data; what truly demonstrates the weight of Muse Spark Open are several practical use cases.

The first use case released by Meta is modifying iOS applications.

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

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

Traditional video editing tools mostly perform basic tasks like automatic subtitling and simple transitions, while Muse Spark Open makes subjective decisions like "which scenes are the best". It can view frame by frame, understand the content, make judgments, and then directly operate professional editing software to produce the final product.

If this process works smoothly, it will directly enhance efficiency for short video production, content creators, and media organizations.

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

Specifically, the model can directly operate a browser, browse e-commerce websites, compare products, and complete purchases. This "AI helps you shop" scenario has been discussed for a long time but has always been limited by the model's understanding of web pages and operational stability. Spark Open's multimodal understanding and tool-calling capabilities have lowered the threshold for such applications.

Sophia Yang, the developer relations director at Fireworks AI, demonstrated a more dazzling application—using Spark Open to generate code frame by frame to create a complete interactive animated webpage.

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

Behind these use cases is Meta's comprehensive layout in open-source models.

In April this year, the first generation of Muse Spark was released, followed by the launch of Glimmer with 30 billion parameters and Apache 2.0 protocol in August, and then the higher-end Muse Wonder and Wonder Full in early October. Now, Spark Open fills the gap in open-source weights, forming a complete product line from small local models to large closed-source models.

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

Internal deployment, research, and non-model service commercial applications are not restricted. This design is very precise—it's fine for internal use, and it's okay to use it to create your own products, but if you want to sell it directly as an API, then revenue sharing needs to be discussed.

This content is provided for general informational purposes only and doesn't constitute financial, investment, legal, or tax advice. Any events, rewards, online promotions, or related information mentioned herein should not be considered a recommendation, solicitation, or invitation to purchase, sell, trade, or otherwise deal in any crypto assets. Crypto assets are highly volatile and may result in loss. The availability of WEEX services, products, and related events may vary by region. You are responsible for ensuring that your participation is in accordance with applicable local laws and regulations.