Muse-Glimmer-30B

Hardware Requirements

Estimated VRAM per quantization at 4,096 tokens context

Only datacenter GPUs — FP16 needs 66.8 GB. For consumer hardware, choose the INT4 build with a multi-GPU setup.

Consumer GPUs that fit this model with a single card (INT4):

Architecture

Training Pipeline

1. 0
   other

   Distillation from Muse Spark

   Distillation from Muse Spark - the larger parent model. Muse Glimmer is purpose-built for autonomous agentic tasks on consumer hardware, inheriting capabilities from Muse Spark while being optimized for local deployment with quantization support.

2. 1
   pretraining

   Pretraining

   Muse Glimmer is a 30-billion-parameter causal language model with a dedicated perception encoder, distilled from Muse Spark. Pretrained on multimodal content sourced from publicly available data, data provided by third parties and information from Meta's products and services. Training data curated and enriched by external vendor networks and Meta personnel. Knowledge cutoff: January 4, 2026. Trained on data from more than 100 languages.

3. 2
   sft

   Safety SFT and Agentic Fine-Tuning

   Curated examples demonstrating correct safety behavior, including agentic safety scenarios covering tool-use boundaries, prompt injection resistance, and permission handling. Fine-tuned for autonomous agentic tasks including multi-step planning, sequential tool invocation, failure recovery, and long-horizon task execution.

4. 3
   rl

   Safety RL and Appropriate Information Flows

   Reinforcement learning with safety-specific reward signals that penalize policy violations while rewarding helpful responses to legitimate requests. Principles of data sensitivity recognition, minimization, and local-first execution embedded directly into model weights through dedicated synthetic training data.

Training & Evaluation Datasets

Linked Resources