Unsloth Releases GGUF Quantizations of Meta's Muse Glimmer 30B Agentic Model
Unsloth has published GGUF quantizations of Muse Glimmer-30B, a dense 29.6B-parameter causal transformer with a dedicated perception encoder, attributed to Meta Superintelligence Lab in the model card. The model targets autonomous agentic tasks on consumer hardware with a 131,072-token context window and 4-bit quantization under 20GB.
Unsloth has released GGUF quantized versions of Muse Glimmer-30B, a dense causal transformer with a dedicated perception encoder designed for autonomous agentic tasks on consumer hardware. According to the model card hosted on Hugging Face, the model is authored by "Meta Superintelligence Lab," carries an Apache 2.0 license, and lists a release date of August 2026 — details that have not been independently confirmed and should be treated as claims from the source repository rather than a verified Meta announcement.
Architecture and Specs
Muse Glimmer totals approximately 29.6 billion parameters, including a ~1.8B-parameter ViT-G/14 perception encoder for image understanding. The language backbone uses 52 layers, a 6,656 hidden dimension, grouped-query attention at a 16:1 ratio (32 query heads / 2 KV heads), SwiGLU feed-forward layers with a 19,968 intermediate dimension, and RoPE position encoding with theta 500,000 applied to local attention layers. Context length is listed at 131,072+ tokens. The tokenizer uses a 202,048-entry vocabulary. Training data has a knowledge cutoff of January 4, 2026, per the model card. The model is described as distilled from a larger model called "Muse Spark."
Local Deployment Focus
Unsloth applied its Dynamic 2.0 quantization to compress the model to roughly 4-bit precision, shrinking it to under 20GB — small enough to run alongside a KV cache and the perception encoder within a 24GB or 32GB VRAM budget. Unsloth claims this compression introduces only 0.2%–1.0% degradation across an internal average of 15 benchmarks, comparing a full-precision 64GB configuration against 32GB and 24GB quantized variants.
The release also ships a companion "drafter" model based on DFlash for speculative decoding, proposing blocks of 16 tokens per forward pass for verification by the main model. Reported throughput gains include 74.9 to 233.4 tokens/second (3.1x) on an Nvidia RTX 5090, and smaller gains of 1.5x–1.8x on Apple M4 Max and M5 Max chips, according to the source.
Benchmark Claims
The model card reports Muse Glimmer scoring 75.5 on MCP-Atlas, 74.6 on DeepSearch QA, 76.0 on SWE-Bench Verified, 94.7 on AIME 2026, and 83.5 on GPQA Diamond — positioned against comparison models "Gemma4-31B" and "Qwen3.6-27B" in thinking mode, neither of which currently exists as a confirmed public release. These comparisons, along with all benchmark figures, come directly from the uploaded model card and have not been independently verified.
Pricing and Access
As an open-weight model distributed via Hugging Face, there is no per-token API pricing; compute cost is limited to local hardware. Pricing not yet disclosed for any hosted inference offering.
What This Means
The unusual details in this listing — a named lab ("Meta Superintelligence Lab"), a forward-dated release month of August 2026, and benchmark comparisons against models that are not yet publicly confirmed — warrant caution. If genuine, Muse Glimmer represents a serious push toward compact, locally-runnable agentic models with multimodal input and speculative decoding built in. Until Meta or another primary source confirms authorship and release timing, this should be treated as an early or leaked model card rather than a fully verified product launch.
Related Articles
Fish Audio releases Drama 3, a 35B-parameter MoE text-to-speech model with 16-speaker support
Fish Audio has released Drama 3, a text-to-speech model with 35B total parameters and 3B active per token, built for multi-character, long-form audio. It supports up to 16 speakers per text, voice cloning, and streaming. Fish Audio reports first audio in under 100 ms on a single H200.
OpenBMB releases MiniCPM-V 4.7 1B, an on-device vision model with 4x/16x visual token compression
OpenBMB has released MiniCPM-V 4.7 1B, a 1B-parameter multimodal model for single-image, multi-image and video understanding on phones, tablets and local hardware. OpenBMB claims it substantially outperforms Qwen3.5 0.8B in overall score and inference efficiency in Instruct mode. Benchmark figures, context window and training cutoff are not given in the available model card text.
Meta launches Muse Wonder agent model, gating it behind $20 and $100 monthly Muse app plans
Meta released Muse Wonder, a model it claims is its most advanced for powering digital assistants. Access requires a $20 or $100 monthly Muse app subscription, and a larger Muse Wonder Full variant is limited to Muse Code subscribers. API pricing has not been disclosed.
Microsoft releases FrogNano-4B, an Apache 2.0 coding agent trained with RL on 1,500 synthetic tasks
Microsoft has released FrogNano-4B-2609, a repository-level coding agent derived from Qwen3.5-4B and published under Apache 2.0 with open weights. Microsoft says it was post-trained only with reinforcement learning on about 1,500 synthetic software-engineering tasks, with no stronger-model trajectories. It is evaluated at roughly 131K tokens of context.
Comments
Loading...