Summary
Meta's Superintelligence Labs open-sourced Muse Glimmer on August 10 — a 30B-parameter agentic model under the Apache 2.0 license, quantized to under 20GB and runnable on a single consumer GPU. Distilled from the flagship Muse Spark, it targets always-on local agent workflows: tool calling, multi-step reasoning, failure recovery, and multimodal input. On agentic benchmarks like SWE-Bench and MCP-Atlas, Meta says it beats same-size rivals Gemma4-31B and Qwen3.6-27B. With DFlash speculative decoding, decode speed improves 3.1x on an RTX 5090.
Body
Meta open-sourced Muse Glimmer yesterday (August 10) through Superintelligence Labs: 30B parameters, Apache 2.0 license, weights live on HuggingFace. The positioning is clear: an always-on local agent model, small enough to run on one consumer GPU, on a Mac or PC, no network required.
Specs first. Muse Glimmer is a 30B dense model with a dedicated perception encoder, accepting interleaved text and images: screenshots, charts, documents. Training follows a distillation path: pre-training uses logit distillation on outputs from the teacher model Muse Spark, mid-training adds longer-context, agent-heavy data, and post-training combines SFT with on-policy distillation and reinforcement learning. In short, it's an attempt to compress flagship agentic capability into an edge-sized model.
The highlighted capabilities: end-to-end task completion (DeepSearch QA, MCP-Atlas, τ-Bench, SWE-Bench), reliable tool calling, coherent multi-step reasoning, diagnosing and retrying failed tool calls, controllable reasoning effort, and support for 100+ languages. It's evaluated against same-size Gemma4-31B and Qwen3.6-27B, with Meta claiming stronger results on agentic benchmarks. Exact numbers live in the evaluation report. The community notes this size class is back in fashion for dense models.
Running locally is the core selling point. At full precision a 30B model needs 55GB+ of memory; 4-bit quantization shrinks the language model to ~17GB, leaving room for the KV cache, perception encoder, and speculative decoding drafter within a 24GB or 32GB envelope. Meta validated minimal degradation on agentic tasks after quantization. For speed, a lightweight DFlash-based drafter proposes blocks of tokens that the main model verifies in parallel — measured 3.1x faster decode on RTX 5090, 1.8x on M5 Max, 1.5x on M4 Max.
The ecosystem story is complete. Optimized integrations for llama.cpp, MLX, and ExecuTorch land in the coming days; vLLM and SGLang serve it at scale; Ollama, LM Studio, and Unsloth are onboard for local use; Together AI, Fireworks AI, and OpenRouter host it. AMD, Arm, Dell, Intel, and NVIDIA are hardware partners. Fine-tuning is supported via PyTorch's TorchTitan. It also works with agent orchestration frameworks like OpenClaw.
Community reaction has been loud. The HN thread passed 1000 points within a day; the top discussions are how Muse Glimmer will compare with Qwen3.8 27B releasing this week, and Meta's tease of an open-weight Muse Spark 1.2. Some read the timing as a direct play for ecosystem position in the open-source size class against Chinese labs. Others note that dense 30B is hot again.
For developers, Muse Glimmer lowers the bar for running agents on-device. Previously, local agents meant settling for smaller models or accepting the latency and privacy tradeoffs of cloud calls. Now a 17GB quantized build fits in VRAM and runs persistently, keeping private data on the device. It's still a 30B model, and the gap vs cloud frontier models on complex reasoning is real, but for tool calling, file organization, and automation, where latency and privacy matter, it's a solid option.
Weights: meta-models/Muse-Glimmer-30B on HuggingFace, developer docs on the Meta AI Developer Center. Within a day of release, the model card already had 700+ likes, with GGUF and ExecuTorch builds published alongside.