Meta unveiled Muse Glimmer, a 30 billion-parameter large language model, on Monday, marking the company's first open weights model release in more than a year. The model is distilled from Meta's larger proprietary Muse Spark system and arrives under a highly permissive Apache 2.0 license that lets businesses deploy, use, and modify it freely. The launch signals Meta's return to open source AI development after the company appeared to abandon that strategy following Llama 4's poor reception and a restructuring of its AI division earlier this year.
Muse Glimmer is designed for local AI inference tasks including agents, code assistants, and applications requiring multi-modal tool use and function calling. At its native BF16 precision, the model requires around 60 GB of memory and fits comfortably into a single Nvidia RTX Pro 6000 or AMD MI350P. Quantized to 4-bit precision, the weights shrink to just under 16 GB, small enough to run on consumer graphics cards like an RTX 30/4090 or RX 7900 XT/XTX. On an RTX 5090 with 1.8 TB/s of memory bandwidth, users can expect between 75 and 233 tokens per second, with the higher end relying on speculative decoding that uses a small draft model to predict outputs. An M5 Max MacBook Pro delivers 26.2 to 57.8 tokens per second, while typical Windows notebooks without dedicated graphics cards produce around 6 to 14 tokens per second.
According to Meta's benchmark figures, Glimmer outperforms Google's comparably sized Gemma 4 31B in most scenarios and trades blows with Alibaba's Qwen 3.6-27B. The company positions the model as ideal for small-to-medium sized enterprises or enthusiasts, competing with similarly-sized offerings from Alibaba and Google rather than larger frontier models. Early support has begun appearing on popular local AI inference platforms including Llama.cpp, Ollama, and Unsloth, with optimized implementations expected over the coming weeks. Meta's Superintelligence chief Alexandr Wang committed on X to releasing an open weights version of Muse Spark 1.2 "soon," though both Kimi K3 and Qwen 3.8-Max currently outperform Spark according to Artificial Analysis' intelligence index.
The release comes as American tech companies confront a crisis over the proliferation of Chinese open weights models and calls for domestically-created alternatives. Meta built its reputation on open weights development starting with the Llama series in 2023, but critics questioned CEO Mark Zuckerberg's commitment to open source AI after Llama 4 flopped. At 30 billion parameters, Glimmer is too small to compete with Moonshot AI's Kimi K3, Alibaba's Qwen 3.8-Max, DeepSeek V4 Flash, or other Chinese models that have dominated AI news cycles in recent months. If Meta's forthcoming open weights Spark model delivers similar performance using fewer parameters, it may prove more competitive from a cost per token standpoint, though the actual size of Spark remains unknown. Developers eager for another fine-tuning base won't have to wait long, and enterprises get a permissive license that removes deployment friction at a time when model sovereignty concerns are rising.
The model's weights are available for download on Hugging Face and through local inference platforms like Ollama and LM Studio. Glimmer uses all 30 billion parameters to generate each token, making it slower than recent models like Qwen 3.6-35B-A3B or Gemma 4 26B-A4B that use just 3 to 4 billion parameters, meaning performance depends heavily on memory bandwidth. Testing on a DGX Spark in Unsloth Studio yielded around 12.2 tokens per second, though DSpark support doesn't appear to have been implemented yet. Meta's pivot back to open weights suggests the company recognizes that maintaining influence in AI infrastructure requires feeding the developer ecosystem, not just chasing proprietary model leaderboards. Whether this strategy can rebuild the goodwill lost during the Llama 4 stumble remains an open question that will likely be answered by how quickly fine-tuners and enterprises adopt Glimmer and the eventual Spark release.