88Z AI, 16 October — Meta has released Muse Spark, an open-source native multimodal model featuring a mixture-of-experts architecture with 816 billion total parameters and 42 billion active ones. Supporting a million-token context, it targets coding and agents. While topping six benchmarks and claiming to be the strongest US open-weight model, it trails Chinese counterparts like Kimi, Qwen, and Zhipu in software development.
At a glance
- Muse Spark features 816B total parameters and a million-token context, focusing on coding and agents.
- Despite topping six benchmarks, it trails Chinese models like Kimi and Qwen in software development.
The update
Meta has officially open-sourced its native multimodal large model, Muse Spark. Built on a mixture-of-experts architecture, the model features 816 billion total parameters with 42 billion active ones, and supports a context window of one million tokens. Its primary applications focus on code generation and agent tasks. Meta claims that Muse Spark is the strongest open-weight model in the United States, achieving first place in six benchmark tests. However, in specific areas such as software development, the model still lags behind Chinese open-source counterparts like Kimi, Qwen, and Zhipu, indicating a gap in performance for these particular tasks compared to top Chinese models.
Sources & editorial note
This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.
- 财联社Source published
Source dates above refer to the original announcements; the date under the headline is this article’s publication on 88Z AI.