[DigitalToday reporter Chi-gyu Hwang (황치규)] As U.S. big tech companies enter the market for open-weight AI models long led by Chinese firms, potential shifts in the competitive landscape have become a key point to watch.

Meta unveiled an open-weight model called "Muse Spark Open" with 816 billion parameters.

VentureBeat reported on Oct. 15 that Meta described Muse Spark Open as the best among U.S. open-weight models released so far.

Muse Spark Open adopts a mixture-of-experts (MoE) architecture. It activates only 42 billion parameters when processing a single token. It supports up to 1 million tokens of context and takes text, images, video and audio as input. Output is limited to text.

Meta also presented a setup for running Muse Spark Open locally by linking two Mac Studio systems with 512 GB of unified memory. Apple is expected to release a 512 GB memory product in late October, and the price has not been disclosed.

The cost burden is large. A Mac Studio with the required M5 Ultra chip and 256 GB of memory currently costs at least $10,799.

Meta said Muse Spark Open outperformed rival models in evaluations using JobBench and MCP-Atlas tools. It trailed Moonshot AI's Kimi K3 and GLM 5.3 in terminal coding and long-term software engineering tasks. Meta put its Artificial Analysis Intelligence Index score at 39. Six Chinese models scored higher, including Xiaomi's MiMo-V2.6-Pro at 46.

Chinese AI companies are leading the competition in open-weight models. Hugging Face, an open-source AI model sharing platform, said a March assessment found Chinese models accounted for 41 percent of downloads over the past year, surpassing U.S. models.

ㆍ[Tech Inside] Everything about open-weight models shaking up the AI landscape

The Muse Spark Open licence allows execution, modification, fine-tuning and redistribution. However, companies with annual revenue exceeding $20 million, including affiliates, must sign a separate contract with Meta to use the model for service businesses. Internal distribution, research and use within a company's own products are exceptions.

AI cloud company Together AI set API pricing at $1.25 per 1 million input tokens and $4.25 per 1 million output tokens.

Nvidia is also taking aggressive steps to expand its presence in the open-weight AI model market. It is considering additional investment in open-source AI startup ReflectionAI or even acquiring the company outright.

Negotiations are at an early stage, the Financial Times reported. The structure of a deal could take several forms. An "acqhire" is also under consideration, in which Nvidia would avoid an acquisition by hiring staff and buying only technology usage rights, given that this could help avoid regulatory scrutiny.

ㆍNvidia makes increasingly bold bets on open-weight AI, considers acquisition of ReflectionAI