Meta Unveils Muse Spark Open
On Thursday, Meta unveiled a new AI model with open weights that it says matches — and sometimes beats — comparable offerings from top Chinese firms. According to Alexandr Wang, Meta's chief of AI, Muse Spark Open stands as the strongest US open-weight model released so far, a claim he made in a post on X. With open-weight systems, as opposed to closed ones like Anthropic's Claude or OpenAI's ChatGPT, developers are able to obtain and alter the model's parameters, referred to as weights and biases.[S1][S2][S3][S4][S5][S6]
Over the last few years, Chinese companies have become the leading force in the open-weight AI model sector. Models such as Alibaba's Qwen, Moonshot AI's Kimi K3 and Z.ai's GLM have been regarded as the top open-weight contenders worldwide, yet Meta asserts that its newly released open model now performs at a comparable level. In a blogpost, Meta stated that Muse Spark Open, designed for coding and agentic tasks, performs strongly on long-horizon work, intricate software engineering and multi-step reasoning, and that robust safeguards were built into it.[S1]
Optimized for Apple Chips
Meta also said Muse Spark Open has been tuned to operate on Apple's high-end M5 Ultra chips, pointing to an uncommon collaboration between the two companies. Running Muse Spark Open demands a substantial amount of computing power. As Meta explained, users can turn two Mac Studios equipped with M5 Ultra chips into a desktop supercomputer capable of running the model.[S1][S7][S8]
This release follows only days after Meta introduced its latest closed model, which the company said could match the most capable iterations of ChatGPT and Claude. Meta's AI strategy originally centered on open models, though Zuckerberg has more recently steered the company toward taking on OpenAI and Anthropic.[S1]
