Muse Glimmer: Meta's 30B agent fits under 20GB, cloud optional

How much useful AI can you fit on a machine you own? Meta's answer is Muse Glimmer, a 30-billion-parameter agentic model released under Apache 2.0. Roughly 4-bit quantization brings the weights under 20 GB, leaving room inside a 24 GB or 32 GB consumer GPU for the model, its working memory, the perception encoder and a small drafter that speeds up decoding. Meta Superintelligence Labs positions it for always-on local agents, function calling, local coding and LLM-as-a-judge evaluation, with or without an internet connection.

Speed has always been the catch with local models. A 30B model grinding out one token at a time turns long reasoning chains and multi-step tool calls into a waiting game, so Meta paired Glimmer with DFlash, a small companion network that proposes whole blocks of tokens for the main model to verify in parallel. Meta measured 3.1x faster decoding on an RTX 5090, 1.8x on an M5 Max and 1.5x on an M4 Max, with output quality unchanged.

What fits in a 24 GB envelope

At full precision, the language model alone needs more than 55 GB, far beyond any consumer GPU. The quantized weights come in under 20 GB, and Meta says the remaining headroom covers the model's working memory, the perception encoder for images and the drafter at the same time. The company says the compression costs minimal to no degradation on agentic tasks. Here is the sizing and speed picture Meta publishes.

Built for agents that fail and retry

The agentic training shows up in concrete capacities: schematized tool calls, multi-step reasoning over long horizons, and failure recovery, where the model diagnoses a failed tool call and retries instead of halting. A dedicated perception encoder accepts interleaved text and images, so an agent can work from screenshots, charts or documents mid-conversation. Different reasoning strengths let developers trade quality against speed, and training data spans more than 100 languages.

On end-to-end agentic benchmarks such as DeepSearch QA, MCP-Atlas, tau-Bench and SWE-Bench, which measure finishing a task from start to finish inside a scaffold, Meta says Muse Glimmer posts strong success rates for its size class. It works with open orchestration frameworks including OpenClaw.

The Muse family goes local

Glimmer did not start from scratch. Meta says it distilled the model from Muse Spark's outputs using logit distillation, kept a similar data mix to its teacher, then ran mid-training on longer, agent-heavy data and post-training with supervised fine-tuning, on-policy distillation and reinforcement learning. The release went through Meta's Advanced AI Scaling Framework before the weights were published.

That connects Glimmer to a family Meta has been assembling fast. Muse Spark 1.2 is the model behind Muse Code, the terminal coding agent Meta shipped in beta for macOS and Linux, with a replay-exact event log and persistent background subagents that keep working without a human at the keyboard. Glimmer takes that reasoning and sizes it for a laptop that might be offline. llama.cpp, MLX and ExecuTorch support land in the coming days; Ollama, LM Studio and Unsloth line up for easy local installs. vLLM and SGLang cover serving at scale, hosted options come via Together AI, Fireworks AI and OpenRouter, and developers can customize the weights further with PyTorch's TorchTitan. Meta lists AMD, Arm, Dell, Intel and NVIDIA as hardware partners. The weights are up on Hugging Face now.

Open weights, owned hardware

The strategic bet is in the license. Apache 2.0 means anyone can fork Muse Glimmer, retrain it or ship it inside a product without paying Meta, the same playbook the company has run for years in research, now pointed at agents. It is a familiar gambit for the lab spending billions on the bet that open weights win. Meta compares the model against Gemma4-31B and Qwen3.6-27B and says it performs strongly for its size class on several widely used benchmarks, though the launch post includes no scores of its own and points to a separate report for the details.

Meta is not alone in chasing on-device agents. Liquid AI made a similar pitch in August with LFM2.5-2.6B, a 2.6-billion-parameter line built to run almost anywhere, in the browser included. Glimmer targets a heavier class of agentic work than Liquid's small models, on the same kind of hardware people already own, and it ships with documentation and setup guidance on Meta's AI Developer Center (dev.meta.ai). The open-weight push also reaches the other end of the size scale, as Alibaba is about to publish Qwen3.8-Max.

The open question is verification. Meta's benchmarks come from Meta, and independent runs on home GPUs will settle how close the experience comes to the charts. The direction is clear either way: agents sized for hardware people already own, weights anyone can fork, and inference treated as a location problem rather than a cloud default.