Meta Launches Muse Glimmer 30B for Local AI as Chinese Models Dominate OpenRouter Usage
Meta is returning aggressively to the open weight AI market with Muse Glimmer, a 30B parameter model designed specifically for local agentic workloads that can run entirely on a single consumer GPU. The release arrives as Chinese AI models continue building momentum across open model platforms, with OpenRouter data showing Chinese models leading token usage for 15 consecutive weeks.
According to Meta AI Research, Muse Glimmer is optimized for persistent local agents, coding, function calling, multimodal reasoning, long context workflows, and automated task execution. The model was distilled from the substantially larger Muse Spark model using logit distillation during training, while separate quantization techniques reduce the final model memory requirements. Meta compresses Muse Glimmer to approximately 4 bit precision, bringing its language model weights below 20GB compared with more than 55GB at full precision. This leaves enough memory for the KV cache, image perception encoder, and speculative decoding system within a 24GB or 32GB memory envelope.
Introducing Muse Glimmer, an open-weight 30B-parameter model optimized for local, always-on agent workflows.
— AI at Meta (@AIatMeta) August 10, 2026
Muse Glimmer delivers strong performance on key agentic use cases and benchmarks compared with leading models in its size category, and is designed to run entirely on… pic.twitter.com/mI4z91GPnE
That distinction matters because distillation and quantization perform different roles. Distillation transfers capabilities from the larger Muse Spark teacher into the smaller Muse Glimmer model during training, while quantization reduces the precision and memory footprint of the finished model for local inference. Meta says the compressed versions introduce minimal or no degradation across its agentic evaluations.
Muse Glimmer also introduces DFlash speculative decoding to improve response speed. The lightweight companion model predicts blocks of 16 tokens in a single forward pass before Muse Glimmer verifies the proposed output in parallel. Meta reports that its 17GB quantized configuration increases generation performance on an NVIDIA GeForce RTX 5090 from 74.9 tokens per second to 233.4 tokens per second, representing a 3.1x improvement. Apple M5 Max performance increases from 26.6 to 50.2 tokens per second, while the M4 Max improves from 23.7 to 37.8 tokens per second.
The model includes multimodal text and image processing, support for more than 100 languages, failure recovery, controllable reasoning effort, tool usage, and a 131,072 token sequence length. Meta is releasing the weights under the Apache 2.0 license and says integrations with llama.cpp, MLX, ExecuTorch, Ollama, LM Studio, Unsloth, vLLM, SGLang, OpenRouter, and other AI platforms are being supported. AMD has already demonstrated Muse Glimmer running locally on Ryzen AI Max systems and a single Radeon AI PRO R9700, reporting up to 53 tokens per second on the Radeon configuration with DFlash enabled.
The timing is particularly significant. OpenRouter processed approximately 69 trillion tokens during the week beginning August 3, with Chinese models accounting for 34.25 trillion compared with approximately 9.17 trillion from US models. This represents the 15th consecutive week in which Chinese models led that specific OpenRouter usage measurement. It should not be interpreted as China controlling all global AI usage, but it provides a strong indication of how rapidly models from DeepSeek, Alibaba, Moonshot, and other Chinese developers are gaining adoption within the open model ecosystem.
The competition is increasingly centered on efficiency rather than parameter count alone. DeepSeek has already demonstrated this direction with DeepSeek V4 and its focus on lower inference cost and memory requirements, while Muse Glimmer gives Meta a new answer focused specifically on capable AI agents that can operate locally without requiring continuous access to expensive cloud infrastructure.
Muse Glimmer may be more strategically important for Meta than another massive frontier model. Running a capable 30B agentic model inside a single 24GB GPU changes the economics of private AI, particularly for developers, gamers, creators, and businesses that want local automation without recurring cloud token costs.
Chinese developers have gained significant traction by aggressively optimizing model efficiency and pricing. Meta appears to be responding with the same battlefield in mind. The next phase of the open AI competition may not be decided by who builds the largest model, but by who delivers the strongest reasoning, agentic performance, privacy, and tokens per second on hardware people can actually own.
Would you run a 30B AI agent locally on your gaming PC if your GPU had enough VRAM, or do you still prefer cloud based AI models?
