.png){kind=logo}
Local AI: Glimmer brings advanced agent capabilities to consumer-class hardware with sub-20 GB quantized versions.
Strong performance: The model delivers competitive results across agent, coding, reasoning, and vision benchmarks.
Fast inference: DFlash raises RTX 5090 performance from 74.9 to 233.4 tokens per second.
Meta has released Muse Glimmer, a new open-weight AI model built for agent tasks, tool use, code work, image input, and long tasks. Meta released the model in August 2026 under the Apache 2.0 license. The model has about 29.6 billion parameters and uses a dense Transformer design. A 1.8 billion parameter vision encoder gives Glimmer the ability to read images, screens, charts, and documents. The model also supports more than 100 languages and has a 131,072-token context window. Its knowledge cutoff sits at January 4, 2026.
Glimmer targets a different need from a normal chatbot. The model aims to act as a local AI agent that can plan tasks, call tools, write code, inspect visual data, recover from errors, and handle several steps in one task. Meta created Glimmer from Muse Spark through a distillation process. Muse Spark remains the larger model, while Glimmer brings part of that capability into a much smaller package. Meta also plans an open-weight release of Muse Spark, which could make the Muse family far more important for the open AI market.
The strongest part of Glimmer may sit in its hardware needs. Meta says a 4-bit version can fit below 20 GB. A 17 GB K-Quant version can fit inside a 24 GB memory setup, while a 32 GB setup can use a larger K-Quant Dynamic version. Meta reports only 1.0% average score loss for the 17 GB version across 15 common tests. The 32 GB version shows just 0.2% average loss. These figures make Glimmer a strong fit for high-end consumer GPUs and some modern AI PCs.
Glimmer also uses a local attention pattern with two local layers followed by a global layer. Its local layers use a 2,048-token window. The model has 52 layers, a hidden size of 6,656, 32 query heads, and two key-value heads. This design helps Glimmer handle long context while keeping memory use under better control.
Also Read - 29 States vs Meta: The Trial that Could Change Social Media
Meta pairs Glimmer with DFlash, a small speculative decoding model. DFlash can propose blocks of 16 tokens at once, then the main model checks those tokens in parallel. Meta reports 74.9 tokens per second without DFlash on an RTX 5090 and 233.4 tokens per second with DFlash. That result gives a 3.1× speed increase. An Apple M4 Max rises from 23.7 to 37.8 tokens per second, while an Apple M5 Max rises from 26.6 to 50.2 tokens per second.
Real local tests show lower results on older hardware, which is normal. One RTX 3090 test reported about 34.6 tokens per second without speculation and 62.6 with DFlash at about 2K context. Larger context sizes can reduce speed. Such results show why the 233.4 figure should stay tied to Meta's RTX 5090 test rather than serve as a universal Glimmer speed figure.
Glimmer shows some of its best results on agent tests. Meta reports a 75.5 score on MCP Atlas, 74.6 on DeepSearch QA, 23.5 on τ3-Banking, 47.6 on WildClawBench, and 43.3 on Gaia2. On several of these tests, Glimmer scores above Gemma 4 31B and Qwen 3.6 27B. Yet Qwen 3.6 27B remains ahead on SkillsBench and OSWorld-Verified, with scores of 46.6 and 75.6 against Glimmer's 44.3 and 65.9.
Glimmer also shows strong code results, though the model does not lead every test. SWE-Bench Pro gives Glimmer 51.2, ahead of Qwen 3.6 27B at 50.2 and Gemma 4 31B at 36.9. SWE-Bench Verified gives Glimmer 76.0, while Qwen reaches 77.2. TerminalBench 2.1 shows a larger gap, with Qwen at 60.7 and Glimmer at 51.7. SciCode gives Glimmer 43.6, just above Gemma at 43.4 and Qwen at 39.8.
General tests show the same mixed picture. Glimmer scores 94.7 on AIME 2026, 83.5 on GPQA Diamond, 22.0 on HLE Text, and 80.0 on AA-LCR. Qwen reaches 94.1 on AIME 2026 and 84.2 on GPQA Diamond, while Gemma reaches 85.7 on GPQA Diamond. These figures place Glimmer among the strongest models in its size class, but they do not support a claim that Glimmer wins every test.
Glimmer also handles visual tasks through its 1.8 billion parameter vision encoder. Meta reports 78.8 on Charxiv Reasoning, 75.4 on ScreenSpot Pro, 75.8 on OmniDocBench 1.5, and 74 on MMMU Pro. The model can also work with tools and agent frameworks such as OpenClaw and Hermes Agent. Support has also reached tools such as llama.cpp, vLLM, SGLang, Ollama, LM Studio, ExecuTorch, and Unsloth.
Meta also tested Glimmer for agent safety. On Siren AgentDojo, Glimmer has a 28.4% attack-success rate and 94.2 utility. Qwen has a 40.3% attack-success rate and 92.7 utility. Meta classifies Glimmer below its Frontier AI threshold, with moderate-or-lower assessments across several risk areas.
Why This Matters?
Muse Glimmer matters as AI moves from simple chat toward independent task execution. Its local hardware support can bring capable agents closer to everyday computers, while open-weight access can give developers more freedom. Its combination of reasoning, vision, tools, and long context also creates strong competition within the 30B model class.
Muse Glimmer stands out less as another 30B chatbot and more as a serious attempt to put agent AI on local hardware. A 29.6B dense model, 131K context, image input, tool use, Apache 2.0 licensing, sub-20 GB quantization, and DFlash speed gains create a rare combination. The benchmark record remains mixed, yet the local agent use case looks unusually strong. Meta's release also signals a wider return to open-weight AI, with a future open-weight Muse Spark release potentially carrying even greater impact.
1. What is Meta Muse Glimmer?
Muse Glimmer is a 29.6B open-weight multimodal AI model from Meta, designed for agentic tasks, tool use, coding, reasoning, and visual inputs.
2. How much memory does Muse Glimmer need?
A 4-bit version can fit below 20 GB, while Meta reports a 17 GB K-Quant version for 24 GB memory systems.
3. Does Muse Glimmer support images?
Yes. Glimmer includes a 1.8B-parameter vision encoder that supports images, screenshots, charts, and documents.
4. How fast is Muse Glimmer?
Meta reports 233.4 tokens per second on an RTX 5090 with DFlash, compared with 74.9 tokens per second without it.
5. Is Muse Glimmer open source?
Meta describes Glimmer as open-weight. The model uses the Apache 2.0 license, which permits broad commercial and development use.
