GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 -2.3%
GPT-6A 820 —
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 586 -0.5%
INKL 531 —
CL-OP46 497 —
CL-OP48 490 -0.2%
GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 -2.3%
GPT-6A 820 —
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 586 -0.5%
INKL 531 —
CL-OP46 497 —
CL-OP48 490 -0.2%
← Back to feed

Meta Releases Muse Glimmer: 30B Open-Weight Agent That Runs on a Single Consumer GPU

Meta Superintelligence Labs has shipped Muse Glimmer, a 30-billion-parameter open-weight model built to run local agent workflows on consumer hardware. Weights are on HuggingFace under Apache 2.0 today. Quantized versions bring the language model to under 20GB, fitting within a 24GB or 32GB memory envelope alongside the perception encoder and KV cache.

This is distinct from Muse Code (the terminal coding agent using cloud-based Muse Spark) and from the broader Muse Spark frontier model line. Muse Glimmer is the local tier — built for always-on personal agents that operate without a network connection.

Architecture and Training

Muse Glimmer was trained in three phases against Muse Spark (the larger teacher):

  • Pre-training: Logit distillation from Muse Spark with a similar data mix.
  • Mid-training: Extended on longer-context, agent-heavy data with richer reasoning traces and organic data.
  • Post-training: Supervised fine-tuning combined with on-policy distillation and RL across general, reasoning, coding, and agentic domains.

The result is a model purpose-built for agent task completion, tool use, multi-step reasoning, and failure recovery — not a shrunk version of a general model. When a tool call fails or returns unexpected output, Muse Glimmer diagnoses and retries rather than halting.

Why This Fits on a Laptop

At full float32 precision, a 30B model needs 55+ GB of memory. Meta applied two techniques to make it run on consumer hardware with acceptable latency:

Quantization: Weights compressed to approximately 4-bit precision via K-Quant methods, producing a 17GB model. This leaves headroom for the KV cache and the DFlash speculative decoding drafter within 24-32GB GPU memory.

DFlash speculative decoding: A lightweight companion network proposes blocks of tokens in parallel; the main model verifies and corrects. The result is materially faster generation than standard autoregressive decoding — with identical output. Meta validated this against agentic benchmarks and reports minimal degradation from the quantization step.

Meta has published token generation benchmarks on RTX 5090, MacBook M4-Max, and M5-Max. Inference is described as fast enough for fluid conversation and real-time agent interaction.

Capabilities

Muse Glimmer was evaluated across the benchmarks that matter for agent deployment:

  • SWE-bench: End-to-end code resolution in scaffolded environments
  • Tau-bench: Retail and telecom multi-turn task completion
  • MCP-Atlas: Tool use across MCP-compatible servers
  • DeepSearch QA: Long-horizon research and retrieval

Meta reports competitive results against Gemma4-31B and Qwen3.6-27B at this size class. Specific benchmark numbers are available in the methodology report at research.meta.ai. Independent evaluations are not yet posted.

Additional capabilities: multimodal input (text and images via a dedicated perception encoder), 100+ language support, controllable reasoning effort (toggle between faster and more deliberate generation), and compatibility with standard orchestration frameworks including llama.cpp, MLX, ExecuTorch, Ollama, LM Studio, vLLM, and SGLang.

What This Is For

Meta’s framing is explicit: agents that manage schedules, draft messages, organise files, and learn user behaviour require access to personal context. That access is easier to grant when the model runs locally and no data leaves the machine. Muse Glimmer is the model for that use case.

This follows a pattern Meta has used before: ship a frontier lab model (Muse Spark) and a smaller local model (Muse Glimmer) that inherits capability through distillation. The local model gets a broader distribution because the use case requires it — always-on agents with deep device access need to be trusted and available offline.

Competitive Position

The 27-35B open-weight tier is now crowded. Direct competitors at the agentic-capable end:

  • Qwen3.6-27B: 77.2% SWE-bench Verified, 579 on Stack Futures agentic index
  • Gemma4-31B: 332 on Stack Futures agentic index (tau-bench only data)
  • Inkling-Small (276B MoE, 12B active): 80.2% SWE-bench, 643 on Stack Futures agentic index

Muse Glimmer targets the same size class as Qwen and Gemma but with an explicit local-deployment-first architecture. Cohere’s North Mini Code (30B MoE) and Bonsai 27B (iPhone-tier) are adjacent products with different primary use cases.

SWE-bench and tau-bench numbers for Muse Glimmer will determine where it ranks on the Stack Futures agentic index once independent evaluations are posted. The current agentic-scores entry will be updated when real numbers are confirmed.

Ecosystem Support

Partners shipping Muse Glimmer support in the coming days: Ollama, LM Studio, Unsloth, Together AI, Fireworks AI, and OpenRouter. Edge frameworks llama.cpp, ExecuTorch, and MLX will have optimized integrations. For production serving, vLLM and SGLang are supported.

Weights are at huggingface.co/meta-models/Muse-Glimmer-30B.