Meta retakes the lead among U.S. open-weight models with Muse Spark Open, which the Interconnects newsletter presents as the most performant model of its kind made in the United States, with a fairly clear margin. According to the author, without this model and Beam, the first Reflection model, more than twenty open-weight models designed by Chinese companies would surpass the best Western models.
- Meta reprend la tête des modèles open-weight américains avec Muse Spark Open.
- Ce modèle de 816 milliards de paramètres dépasse GLM 5.2, mais reste derrière GLM 5.3 et Kimi K3.
- Sa licence restrictive pourrait freiner son adoption par les entreprises.
- Muse Wonder, en version fermée, n’est dépassé que par Opus 5.5 d’Anthropic.
Mistral, for its part, announced Mistral Large 4, which achieves a score of 38, but its weights have not yet been released. Other U.S. open-weight models are reportedly in preparation, according to rumors. The author estimates that the gap between the top U.S. and Chinese models has been reduced by roughly half.
A model with 816 billion parameters
Muse Spark Open has a total of 816 B parameters, including 42 B active ones, and features a context window of 1 M tokens, i.e., the amount of text it can process at once. According to the data sheet published on Hugging Face, it is designed for code and agentic tasks, which involve chaining actions autonomously, and it is natively multimodal. The sheet also notes that this version received an additional post‑training security fine‑tuning and must be deployed with safeguards appropriate to each use case.
According to the author, its performance ranks above GLM 5.2, near DeepSeek V4.1 Flash and behind GLM 5.3 and Kimi K3, whose versions were released on August 18 and July 16 respectively. Some observers feared that this model had been “broken” by the security training after the release of Muse Spark 1.3. The author considers it more likely that it is a secured version of Muse Spark 1.2, which would lose only one point on the Artificial Analysis Intelligence Index. Alex Wang had also indicated, at the launch of Muse Glimmer, a 30 B agentic model, that version 1.2 would be released as open weights.
A license deemed restrictive
According to the author, the specific license of Muse Spark Open could be the main barrier to its adoption. It requires displaying the notice “Built with Muse” on products that use it, and companies generating more than $20 million in monthly revenue across their entire business must obtain a separate license from Meta. It also mandates acceptance of an acceptable‑use policy, which, in the author’s view, distances this model significantly from free software.
These conditions also apply to model releases and its derivatives, creating legal uncertainty around the use of synthetic data. The license, however, permits internal research and internal applications. The author notes that this text has not been tested in court and that many actors might ignore it, with Meta likely pursuing only flagrant violations. Nevertheless, he judges that a complex license represents an unnecessary strategic risk for labs seeking to regain adoption.
An evolving balance of power
A few days before Muse Spark Open, Meta unveiled its closed models Muse Wonder and Muse Wonder Full, which, on traditional benchmarks, are only outperformed by Anthropic’s Opus 5.5 and sit at a level comparable to OpenAI’s GPT‑6‑Astra or Google’s Gemini‑4‑Argon. These announcements come a little over a year after Meta’s pivot to AI, marked by the “Personal Superintelligence” message from its CEO, Mark Zuckerberg.
In his ranking, the author places Anthropic ahead of OpenAI, then Meta and Gemini on equal footing, followed by xAI and the Chinese open-weight models, which themselves rank ahead of the U.S. open-weight models. He estimates that Microsoft could join the leading players if it begins publishing models. Meta is valued at under $2 trillion, and Anthropic is approaching an IPO at a comparable level, according to the author, who wonders what Anthropic’s public valuation will be in November 2027, and then in 2028.
