Meta AI Releases Muse Glimmer: A 30B Open-Weights Agentic Model That Runs on One Consumer GPU
Meta has launched Muse Glimmer, a 30-billion-parameter multimodal mannequin distilled from Muse Spark. It is tuned for always-on native agent workflows, and ships underneath Apache 2.0. A 30B mannequin usually wants over 55 GB of reminiscence at full precision. Meta compresses it to roughly 4-bit, then provides block-level speculative decoding so it solutions quick sufficient to sit down inside an actual agent loop. The end result runs on one client GPU or a Mac, with no community name.
Is it deployable?
Yes, the weights are open underneath Apache 2.0. The Hugging Face collection carries BF16 weights, GGUF k-quants, ExecuTorch builds, and the DFlash drafter. Self-hosting is the day-one path.
- Which firms: Solo builders and startups can run it on one 24 GB GPU or an M4/M5 Max Mac. Mid-market groups get on-prem inference with out a per-token invoice. Regulated enterprises get an air-gappable agent. Meta advises including system-level guardrails slightly than delivery the mannequin as a naked endpoint.
- Industries: Healthcare, authorized, monetary providers, protection and public sector, manufacturing, and discipline service. These are the settings the place information residency, offline operation, or latency rule out a cloud name.
- Applications: Desktop brokers that learn screenshots, coding brokers, and schema-based perform calling. Also doc and chart understanding, artificial information era, and LLM-as-a-judge analysis.
Model and coaching
Muse Glimmer is a dense causal transformer with a devoted notion encoder. Total parameters are roughly 30B, together with the imaginative and prescient tower. Grouped-query consideration makes use of 32 question heads and a couple of KV heads. Attention repeats a [Local, Local, Local, Global] sample with a 2,048 sliding window. RoPE is utilized to native layers solely, with theta 500,000. The imaginative and prescient facet is a ~1.8B ViT-G/14 perception encoder accepting as much as 4,096 visible tokens per picture. Context size is 131,072+, vocabulary is 202,048 tokens, and the information cutoff is January 4, 2026. Input is textual content and picture; output is textual content. Audio will not be supported, and video is processed as particular person frames.
Training ran in three phases:
- Pre-training used logit distillation on Muse Spark’s outputs.
- Mid-training added longer-context, agent-heavy information with richer reasoning traces.
- Post-training mixed supervised fine-tuning with on-policy distillation and reinforcement studying throughout basic, reasoning, coding, and agentic domains.
Fitting 30B onto client {hardware}
At full precision the mannequin wants over 55 GB of reminiscence. Meta compresses weights to roughly 4-bit precision, which brings the language mannequin underneath 20 GB. That leaves headroom inside a 24 GB or 32 GB envelope. The KV cache, notion encoder, and drafter share it. Two quantized builds ship. Okay-Quant-Dynamic targets 32 GB VRAM at 0.2% common degradation. Okay-Quant-17GB targets 24 GB VRAM at 1.0%. Degradation is averaged over accuracy metrics throughout 15 widespread benchmarks.
Generation velocity comes from DFlash, a block-diffusion drafter that predicts 16 tokens in a single ahead move. The primary mannequin verifies the block in parallel. The drafter makes use of 5 layers, sliding-window consideration at 2,048, and 32 question / 8 KV heads. Meta measured Okay-Quant-17GB at batch dimension 1 with grasping decoding. On an RTX 5090, throughput rises from 74.9 to 233.4 tok/s, a 3.1x speedup. Apple M5 Max strikes from 26.6 to 50.2 tok/s, and M4 Max from 23.7 to 37.8 tok/s.
Benchmarks
Meta compares Muse Glimmer towards Gemma4-31B and Qwen3.6-27B in pondering mode. It leads on MCP Atlas at 75.5, towards 54.2 and 62.5. It additionally leads on DeepSearch QA at 74.6, Gaia2 at 43.3, and SWE-Bench Pro at 51.2. Reasoning scores comply with: AIME 2026 at 94.7, IFBench at 77.0, AA-LCR at 80.0. Qwen3.6-27B stays forward on OSWorld-Verified, 75.6 versus 65.9. It additionally leads TerminalBench 2.1 at 60.7 and SWE-Bench Verified at 77.2. The sample is constant. Muse Glimmer wins on agentic orchestration and reasoning. It trails on computer-use and terminal work.
On security, Siren AgentDojo assault success price is 28.4 with utility 94.2. Meta states the mannequin doesn’t meet the Frontier AI definition in its Advanced AI Scaling Framework. It charges chem/bio, cyber, and loss-of-control threat at reasonable or decrease.
Key Takeaways
- 30B open-weights agentic mannequin, Apache 2.0, distilled from Muse Spark.
- 4-bit quantization matches it in 24 GB VRAM at 1.0% degradation.
- DFlash 16-token block hypothesis offers 3.1x decode speedup on RTX 5090.
- Beats each comparators on MCP Atlas, DeepSearch QA, and SWE-Bench Pro.
- Trails Qwen3.6-27B on OSWorld-Verified and TerminalBench 2.1.
Check out the Model weights on HF, Details and Meta Blog. Also, be at liberty to comply with us on Twitter and don’t neglect to hitch our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.
Need to accomplice with us for selling your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar and so forth.? Connect with us
The publish Meta AI Releases Muse Glimmer: A 30B Open-Weights Agentic Model That Runs on One Consumer GPU appeared first on MarkTechPost.
