Muse Glimmer

Released August 2026, Muse Glimmer is Meta's open-weight agentic model with a 131,072-token context window, a dedicated vision encoder, and 4-bit quantization for local use. It's built to replace closed, per-token agent APIs with a model that self-hosts on a single consumer GPU.

Muse Glimmer, Meta's open-weight agentic model released August 2026, leads rival 30B-class open models on agentic benchmarks, scoring 75.5 on MCP-Atlas against Gemma4-31B's 54.2. It runs on a single consumer GPU under 32GB of VRAM, with fully open weights on Hugging Face.

Provider: Meta AI · Family: Muse

More about Meta AI on HokAI

Context window: 131,072 tokens

Input modalities: text, image · Output: text, tool-calls

About Muse Glimmer

Muse Glimmer is Meta's open-weight agentic language model, released August 10, 2026 by Meta Superintelligence Labs. It is a dense causal transformer of roughly 29.6 billion parameters, distilled from Meta's closed frontier model Muse Spark through pre-training logit distillation, longer-context agentic data, and on-policy reinforcement learning. Where Spark is Meta's proprietary flagship, Glimmer is the open sibling built to run local, tool-using agents on a single consumer GPU rather than in a datacenter. On Meta's own published comparison against Google's Gemma4-31B and Alibaba's Qwen3.6-27B, Glimmer leads agentic tasks: 75.5 on MCP-Atlas against 54.2 and 62.5, and 74.6 on DeepSearch QA against 61.7 and 71.1. On AIME 2026 it reaches 94.7%, ahead of both rivals. GPQA Diamond is the one reasoning benchmark where Glimmer trails, at 83.5% against Gemma4-31B's 85.7%. Coding splits by benchmark: Glimmer's 51.2% on SWE-bench Pro beats both, but on SWE-bench Verified, Qwen3.6-27B edges ahead at 77.2% against Glimmer's 76.0%, especially on sustained terminal work and computer control. Glimmer ships with a 131,072-token context window, less than half of Spark's 262,144, with no published needle-in-haystack recall score. Grouped-query attention repeats three local layers, using a 2,048-token sliding window, for every one global layer, trading long-range recall for lower memory use and faster inference. Input is text and image; a 1.8-billion-parameter ViT-G/14 perception encoder handles up to 4,096 visual tokens per image, reading screenshots, charts, and documents. Output is text only, including structured tool calls, and Glimmer was reinforcement-trained to diagnose a failed or unexpected tool result before retrying. Four reasoning levels (low, medium, high, xhigh) trade latency for depth, and Meta claims support for over 100 languages, though published evals focus on English. There's no native audio, and video is handled only as sampled frames, capped at 96. Unlike Spark, Glimmer has no vendor price: weights are free under an open license, and the real cost is hardware. A 4-bit build fits under 20GB of VRAM, down from 55GB or more at full BF16; K-Quant-17GB targets 24GB at a 1.0% quality loss, K-Quant-Dynamic targets 32GB at 0.2%. DFlash, Meta's block-diffusion speculative decoder, lifts RTX 5090 throughput from 74.9 to 233.4 tokens per second, with smaller gains on Apple silicon. Launch-day support covers Transformers, Ollama, LM Studio, vLLM, and SGLang, with llama.cpp, MLX, and ExecuTorch close behind. Self-hosting isn't required: OpenRouter, Together AI, Fireworks AI, and NVIDIA NIM all offer hosted access, and Meta hasn't announced a first-party hosted API of its own for Glimmer. Meta's Preparedness Team rated Glimmer moderate risk or lower on chemical/biological, cyber, and loss-of-control axes; it doesn't meet Meta's Frontier AI bar, since it's less capable than Spark. Evaluation is documented in the shared Muse Spark Safety and Preparedness Report rather than a separate Glimmer report. Training data, cutoff January 4, 2026, blends public web content, licensed data, and Meta product data; the license bars use by anyone under 18. Glimmer suits teams wanting an agentic coding or personal-assistant model on their own hardware for privacy or cost reasons, especially tool-orchestration and document or schedule-management agents, where its lead over rivals is largest. Voice products should look elsewhere given the lack of native audio. Workloads dominated by sustained terminal sessions may prefer Qwen3.6-27B, and anyone needing a longer context window or Meta's frontier reasoning score should use closed Muse Spark instead. Glimmer arrives four months after Spark's April 8, 2026 debut and days after Spark 1.2 shipped on August 5, 2026. MSL head Alexandr Wang said an open-weight Spark 1.2 is planned, positioning Glimmer as the first of a recurring open track alongside closed Spark, not a one-off.

Pricing

Muse Glimmer's own weights are free under an open license; there is no official Meta price. On OpenRouter, the only independently verified third-party host as of August 2026, it costs $0.35 per million input tokens and $1.50 per million output tokens. Self-hosting instead costs only hardware and electricity, starting around 20GB of VRAM for the most compressed quantization.

Key Features

• 29.6B Dense Agentic Model: A dense causal transformer distilled from Meta's closed Muse Spark model through logit distillation and on-policy reinforcement learning, built for multi-step tool-using agents rather than general chat.
• 131K Context Window: Handles up to 131K tokens of input, enough for large codebases or multi-document agent workflows, though it's half of sibling model Spark's window.
• Perception Encoder for Vision: A dedicated 1.8B-parameter ViT-G/14 encoder reads screenshots, charts, and documents at up to 4,096 visual tokens per image.
• DFlash Speculative Decoding: A block-diffusion drafter predicts 16 tokens per forward pass, roughly tripling decode speed on high-end GPUs like the RTX 5090.
• Four Reasoning Effort Levels: Low, medium, high, and xhigh settings let a caller trade latency for deeper multi-step reasoning on a per-request basis.

Pros

• Leads rival open agentic models on Meta's own benchmark suite, with the widest published margin on MCP-Atlas of any model in the comparison.
• Meta's K-Quant-17GB quantization fits 24GB of VRAM at just a 1.0% average quality loss versus full precision, unusually efficient for a model this size.
• DFlash speculative decoding delivers up to a 3.1x throughput speedup on high-end consumer GPUs, with no measured accuracy loss.

Cons

• The one head-to-head loss in Meta's own comparison: a rival open model narrowly wins SWE-bench Verified, and pulls further ahead on sustained terminal or computer-use coding tasks.
• Text and image only: no audio in or out, and video support is capped at 96 sampled frames rather than a dedicated video encoder.
• Context window (131K tokens) is half of closed sibling model Muse Spark's 262K, so very long documents may need chunking.

Benchmarks

• mmmu pro: 74
• aime 2026: 94.7
• mcp atlas: 75.5
• gpqa diamond: 83.5
• tau3 banking: 23.5
• deepsearch qa: 74.6
• swe bench pro: 51.2
• charxiv reasoning: 78.8
• swe bench verified: 76
• artificial analysis intelligence index: 35
• artificial analysis price blended per m: 0

Visit Muse Glimmer Official Page