Meta unveils Muse Glimmer, a lightweight AI model you can run on single PC
Meta says, “Muse Glimmer is optimized for these local use cases,” as it helps run models locally and can help users access “AI anywhere, anytime, with or without an internet connection.”
- Aug 10, 2026,
- Updated Aug 10, 2026 5:32 PM IST
Meta, the social giant, has launched an open agentic model called “Muse Glimmer,” which is small enough to run locally on a Mac or PC. Meta says it's a distilled version of Muse Spark 1.2 and features 30 billion parameters. Despite being compact, it is efficient enough to run on a single graphics processing unit (GPU).
Rather than providing responses, it is designed to perform agentic-like tasks on a user's behalf such as managing schedules, organising files, and potentially handling other multi-step computer tasks.
Meta says, “Muse Glimmer is optimized for these local use cases,” as it helps run models locally and can help users access “AI anywhere, anytime, with or without an internet connection.”
Muse Glimmer: Capabilities and benchmarks
Muse Glimmer is released under a permissive Apache 2.0 license, and it can run on Mac or PC with a single consumer GPU. Meta says its use cases include local agents, function calling, local coding, and LLM-as-a-judge evaluation.
Meta says that the model is trained through a multi-stage process, majorly trained on the outputs of its larger Muse Spark model using logit distillation, then advancing it through a mid-training phase on longer-context, agent-heavy data, and finally post-training that combined supervised fine-tuning with on-policy distillation and reinforcement learning across reasoning, coding, and agentic domains.
According to Meta, Muse Glimmer can perform multi-step reasoning, make tool calls, recover from failed tool calls, and understand both text and images. It is also trained on data spanning more than 100 languages.
Muse Glimmer would require more than 55GB of memory, which is beyond what most consumer GPUs offer. To address this, Meta used quantisation to compress the model's weights to approximately 4-bit precision, bringing the model's size down to under 20GB.
In terms of performance, the company tested the model across benchmarks including DeepSearch QA, MCP-Atlas, τ-Bench, and SWE-Bench, which test capabilities such as coding, tool use, and completing multi-turn tasks. It also compared Muse Glimmer with models including Gemma4 31B and Qwen3.6 27B, revealing that it performs strongly for its size.
Muse Glimmer availability
Muse Glimmer is available via Hugging Face, with support for local deployment platforms. It can also be accessed through platforms like Ollama, LM Studio, and Unsloth, and frameworks such as llama.cpp, ExecuTorch and MLX will also support the model.
The company is additionally working with hardware companies including AMD, Arm, Dell, Intel, and Nvidia to optimise Muse Glimmer across different devices.
For Unparalleled coverage of India's Businesses and Economy – Subscribe to Business Today Magazine
Meta, the social giant, has launched an open agentic model called “Muse Glimmer,” which is small enough to run locally on a Mac or PC. Meta says it's a distilled version of Muse Spark 1.2 and features 30 billion parameters. Despite being compact, it is efficient enough to run on a single graphics processing unit (GPU).
Rather than providing responses, it is designed to perform agentic-like tasks on a user's behalf such as managing schedules, organising files, and potentially handling other multi-step computer tasks.
Meta says, “Muse Glimmer is optimized for these local use cases,” as it helps run models locally and can help users access “AI anywhere, anytime, with or without an internet connection.”
Muse Glimmer: Capabilities and benchmarks
Muse Glimmer is released under a permissive Apache 2.0 license, and it can run on Mac or PC with a single consumer GPU. Meta says its use cases include local agents, function calling, local coding, and LLM-as-a-judge evaluation.
Meta says that the model is trained through a multi-stage process, majorly trained on the outputs of its larger Muse Spark model using logit distillation, then advancing it through a mid-training phase on longer-context, agent-heavy data, and finally post-training that combined supervised fine-tuning with on-policy distillation and reinforcement learning across reasoning, coding, and agentic domains.
According to Meta, Muse Glimmer can perform multi-step reasoning, make tool calls, recover from failed tool calls, and understand both text and images. It is also trained on data spanning more than 100 languages.
Muse Glimmer would require more than 55GB of memory, which is beyond what most consumer GPUs offer. To address this, Meta used quantisation to compress the model's weights to approximately 4-bit precision, bringing the model's size down to under 20GB.
In terms of performance, the company tested the model across benchmarks including DeepSearch QA, MCP-Atlas, τ-Bench, and SWE-Bench, which test capabilities such as coding, tool use, and completing multi-turn tasks. It also compared Muse Glimmer with models including Gemma4 31B and Qwen3.6 27B, revealing that it performs strongly for its size.
Muse Glimmer availability
Muse Glimmer is available via Hugging Face, with support for local deployment platforms. It can also be accessed through platforms like Ollama, LM Studio, and Unsloth, and frameworks such as llama.cpp, ExecuTorch and MLX will also support the model.
The company is additionally working with hardware companies including AMD, Arm, Dell, Intel, and Nvidia to optimise Muse Glimmer across different devices.
For Unparalleled coverage of India's Businesses and Economy – Subscribe to Business Today Magazine