Meta Muse Glimmer, a new agentic model with 30 billion parameters that runs locally

iGuRu.gr > Meta Muse Glimmer, a new agentic model with 30 billion parameters that runs locally

Meta has been steadily growing its Muse family of AI models in recent months. In April, the company first announced Muse Spark, built by the newly formed Meta Superintelligence Labs with promising benchmark results. In June, Meta released Muse Spark 1.1, a major upgrade to Muse Spark with improvements to agentic tasks, computing, scheduling, and multimodal understanding. meta muse glimmer

Earlier this month, Meta released the Muse Code agent powered by Muse Spark 1.2 to take on OpenAI’s Codex and Anthropic’s Claude Code in a highly competitive market. Meta is now targeting AI workloads that can run entirely on consumer hardware.

See more articles from iGuRu.gr when you search for news on Google.

iGuRu.gr preferred source

Meta today announced Muse Glimmer, a new 30 billion-parameter agentic model that is optimized for always-on local agents and can run on Mac or PC. It only requires a consumer GPU. Meta has also released the model code under the permissive Apache 2.0 license.

Meta says that Muse Glimmer supports several features required for typical agentic workloads, such as long-horizon reasoning, reliable tool invocation, error recovery, encoding, and multimodal reasoning. In addition, Glimmer supports over 100 languages and works with agent frameworks such as OpenClaw.

Running a full-precision 30B model locally normally requires over 55GB of memory. However, the quantized 4-bit version reduces the model size to less than 20GB, allowing the full setup to run on 24GB or 32GB of memory. Meta also uses DFlash’s speculative decoding to improve token generation speed. DFlash offers a 3,1x speedup with an Nvidia RTX 5090, 1,8x with an M5 Max, and 1,5x with an M4 Max.

Muse Glimmer is available through Hugging Face , with support for Ollama, LM Studio, llama.cpp, MLX, ExecuTorch, vLLM, and SGLang.