Meta Superintelligence Labs has introduced Muse Glimmer, a 30-billion-parameter open-weights model released under the Apache 2.0 license, specifically engineered for always-on local agentic workflows. Optimized for consumer hardware, it enables on-device tasks such as function calling, coding, and LLM-as-a-judge evaluation without reliance on cloud infrastructure.

Core Methodology

The development of Muse Glimmer focused on balancing high-level agentic capabilities with the memory and compute constraints of local hardware through three distinct training phases:

1. Pre-Training: The model was trained using logit distillation based on the outputs of the larger Muse Spark model, utilizing a similar data mix to ensure knowledge transfer.
2. Mid-Training: The model underwent specialized training on longer-context and agent-heavy datasets, incorporating rich reasoning traces to improve performance on complex workflows.
3. Post-Training: A hybrid approach was employed, combining supervised fine-tuning (SFT) with a mix of on-policy distillation and reinforcement learning (RL) across coding, reasoning, and general agentic domains.

Agentic Capabilities

Muse Glimmer is designed for end-to-end task completion, demonstrated by its strong performance on benchmarks like SWE-Bench and $\tau$-Bench. Key features include:

• Reliable Tool Use: Ability to invoke tools with precise schemas across extended sequences.
• Failure Recovery: An inherent diagnostic ability to handle unexpected tool outputs or errors.
• Multimodal Perception: Integration of a dedicated perception encoder that allows for interleaved text and image processing.
• Controllable Effort: Mechanisms that allow the user to balance reasoning strength against latency requirements.

Inference and Optimization

To maintain performance on consumer devices (e.g., Mac M4-Max or RTX-5090), the model utilizes two primary optimization strategies:

• Quantization: By compressing the 30B model to approximately 4-bit precision, the memory footprint is reduced to under 20 GB. This allows the model, its KV cache, the perception encoder, and the speculative drafter to fit within a 24 GB to 32 GB memory envelope.
• Speculative Decoding: To mitigate the latency of token-by-token generation, Muse Glimmer employs a "drafter" model based on DFlash. The drafter proposes a block of tokens in parallel, which the primary model then verifies. If $P$ is the probability distribution of the main model and $Q$ is the distribution of the drafter, the verification process allows the model to accept $k$ tokens per draft step, significantly accelerating generation while maintaining output parity with standard autoregressive decoding.

The model supports integration with frameworks such as llama.cpp, MLX, ExecuTorch, vLLM, and SGLang, and is further customizable via PyTorch’s TorchTitan feature.