Meta opens Muse Glimmer, a 30B Apache 2.0 agent model for 24GB GPUs

Meta's Muse Glimmer is a 29.6B dense model distilled from Muse Spark, released as Apache 2.0 weights and sized to run agents on one 24GB GPU.

Key points

• Repo meta-models/Muse-Glimmer-30B: about 29.6B dense parameters, roughly 1.8B of them a ViT-G/14 vision encoder, 52 language-model layers, text and image in, text out, 131,072+ token context, knowledge cutoff January 4, 2026, trained on 100+ languages.
• Apache 2.0 applies to the BF16 weights, the two roughly 4-bit K-Quant builds, the DFlash drafter head and the frozen perception encoder. It is a weights-only release: VentureBeat notes Meta has not published training data or training code.
• Meta's own table: 75.5 MCP Atlas, 51.2 SWE-Bench Pro (Gemma 4-31B 36.9, Qwen3.6-27B 50.2), 76.0 SWE-Bench Verified, 65.9 OSWorld-Verified, 51.7 TerminalBench 2.1. That same table has Qwen3.6-27B ahead on SWE-Bench Verified (77.2), OSWorld-Verified (75.6) and TerminalBench 2.1 (60.7).
• Independent check: Artificial Analysis, which had pre-release access, scores Glimmer 35 on its Intelligence Index against 38 for Qwen3.6-27B, 36 for Kimi K2.5 and 30 for Gemma 4 31B; 52% on Terminal-Bench 2.1 versus Qwen's 61%; 953 Elo on GDPval-AA v2 versus Qwen's 1141; but 24% on Tau3-Banking, ahead of Qwen's 17%.
• The 4-bit builds bring the language model under 20GB: K-Quant-17GB targets 24GB cards such as the RTX 3090/4090, K-Quant-Dynamic targets 32GB (RTX 5090, 32GB Macs), and BF16 needs 55GB+. Meta's claim is an average accuracy loss of 0.2% for the Dynamic build and about 1% for the 17GB build over 15 benchmarks.
• Availability: weights on Hugging Face at launch; Ollama, LM Studio, vLLM, SGLang, Together AI, Fireworks AI and OpenRouter in launch week; llama.cpp, MLX and ExecuTorch optimisations promised within days, with Unsloth named as a local-runtime partner. As of September 10, Meta's dev docs describe a llama.cpp path (GGUFs at meta-models/Muse-Glimmer-30B-GGUF, build b10353 or newer, DFlash via --spec-type draft-dflash) and Unsloth ships 2-bit through BF16 GGUFs plus fine-tuning on a 24GB card.
• Built by logit-distilling Muse Spark during pre-training, then mid-training on long-context agent data, then SFT, on-policy distillation and RL. Meta says DFlash speculative decoding lifts decode 3.1x on an RTX 5090 (74.9 to 233.4 tok/s), 1.8x on an M5 Max and 1.5x on an M4 Max. Zuckerberg also said Muse Spark 1.2 weights will be released later.

“can run on 24GB of VRAM without losing agentic reliability.” — Alexandr Wang, Meta chief AI officer, on X, as quoted by VentureBeat

Why it matters

For a small studio, this is the first Meta model in a while that can ship inside a client product under a plain Apache 2.0 licence, and it is sized for a GPU a freelancer already owns rather than a rented H100. On the numbers, though, treat Glimmer as a local-first agent runtime to try on your own repos, not a replacement for a hosted frontier model: Meta's own table has Qwen3.6-27B ahead on three of the four coding and computer-use benchmarks it compares, and Artificial Analysis's independent run puts Glimmer a few points behind Qwen on its index and well behind on Terminal-Bench. The llama.cpp path with DFlash is documented on Meta's dev site now, so the practical question is whether the 17GB quant plus speculative decoding on a 3090 or 4090 is fast and reliable enough for an agent loop you would bill for.

This is an AI-written summary of the reporting credited above, read and edited by Nicholas before publishing. The facts and any quote belong to the original; the wording is ours. Read the original.