FULL STORY

Meta Open-Sources Muse Glimmer: A New Standard for Local AI Agents

Meta released Muse Glimmer, a 30B open-source multimodal model optimized for local agents. Community tests revealed exceptional inference efficiency on consumer hardware.

2026-08-10 ~ 2026-08-12 · 8 episodes · 139 posts

Episode 1 · Meta Open-Sources 30B Muse Glimmer for On-Device Agents (2026-08-10, 108 posts)

Meta's Superintelligence Lab released Muse Glimmer, a 30B-parameter open-source multimodal model optimized for on-device agent workflows on consumer hardware. Under Apache 2.0, it supports 128K+ context. Developer tests show it runs efficiently on minimal resources and demonstrates autonomous deployment, environment adaptation, and long-horizon tool use, signaling a shift from passive generation to active system task execution.

Confirmed

  • Release and ecosystem support: Meta released Muse Glimmer, a 30B dense multimodal model with Gemma-like architecture and GQA. Hugging Face and UnslothAI provided day-zero support; Unsloth uploaded GGUF quantized versions and deployment guides. Meta also teased upcoming weights for Muse Spark 1.2.
  • Autonomous deployment and system optimization: Developer benburtenshaw tested tool-calling abilities; the model self-quantized and deployed to a protected Hugging Face Inference Endpoint, auto-completing health checks and chat completion validation. In advanced tests, it optimized deployment configs for H100 architecture.
  • Low-resource tool calling: UnslothAI demonstrated 100+ consecutive tool calls using 2-bit GGUF quantization, consuming only 14GB RAM.
  • Local run and experience: Developer MaziyarPanahi ran it on Mac Studio at 30 tok/s API speed; benburtenshaw set up a research agent on Hugging Face Hub.
  • Performance and benchmarks: On 24 benchmarks, Muse Glimmer achieved best scores in 12, outperforming Gemma4-31B and Qwen3.6-27B (per kimmonismus). ForsookComparison found its chain-of-thought highly efficient, surpassing popular models of similar size in some scenarios.
  • Hardware requirements and quantization: Official says 24GB VRAM needed (m16); danielhanchen retweeted 18GB VRAM. coder543 tested on RTX 3090 with Q4KXL, DFlash, and mmproj supporting 256k context at 22GB VRAM; ydnar shared optimized configs for RTX 3090.
  • Coding and inference speed: Ok-Shower7286 achieved 280 t/s on local coding tasks with DFlash and llama.cpp; curiousily on M5 Pro with Q4 used 20GB RAM at 17 token/s.

Unconfirmed

  • Whether the model fully supports 256k context on 24GB VRAM: coder543's test shows feasible, but performance may vary with different quantizations.
  • Relative capability vs. Qwen: curiousily thinks coding is inferior to Qwen, but kimmonismus cites benchmarks showing superiority over Qwen3.6-27B; disagreement exists.

Why it matters

  • These tests demonstrate AI models transitioning from passive text generation to active complex system tasks. The model's autonomous deployment, environment adaptation, and hardware optimization provide valuable reference for future AI-driven ops.
  • The ability to run 100+ tool calls on 14GB RAM significantly lowers hardware barriers for on-device complex agents, making persistent local agents feasible.
  • Zuckerberg reiterated Meta's commitment to open-source AI, calling for unrestricted model distillation. eyishazyer notes this reflects Meta's dual strategy: keeping flagship models closed for commercialization while open-sourcing distilled compressed versions.

88 more related posts →

Episode 2 · Meta to Open-Source 30B Muse Glimmer, Tops Open-Source Benchmarks (2026-08-10, 6 posts)

Multiple leaks and benchmark data indicate that Meta is about to release a new open-source large model codenamed Muse Glimmer. The model has 30 billion (30B) parameters, uses the Apache 2.0 license, and its configuration files and benchmark scores have been exposed on Hugging Face and Artificial Analysis. This marks Meta's renewed push into the open-weight model ecosystem after the Llama series.

Confirmed

  • Model parameters and license: According to Artificial Analysis's test page and leaked Hugging Face config files, Muse Glimmer has 30B parameters and uses the Apache 2.0 open-source license.
  • Performance: Benchmark data shows the model (Muse Glimmer high) performs excellently, topping open-source leaderboards.
  • Future plans: Community sources indicate Meta also plans to open-source Muse Spark 1.2. As reported by @haider1, Muse Spark 1.2 matches GPT-5.6 Terra and Opus 4.8 in benchmarks, with API costs only one-third of Terra's.

Unconfirmed

  • Architecture details: Some leaks suggest the model uses knowledge distillation from Muse Spark and has an architecture similar to Gemma 4 (a derivative of Llama 3 architecture), but this remains speculative.

Why it matters

  • Reshaping the open-source landscape: According to early testers, while the model has not yet reached the ceiling of open-source models from Chinese vendors and lags behind closed-source frontier models, it is already the strongest non-Chinese open-source model released globally in the past year. Commentators believe this will be the largest open-weight model from Meta's ecosystem since Llama 3 and Llama 4, potentially reclaiming the top non-Chinese open-source position.

Episode 3 · Leaked Architectures of New 30B to 400B Parameter Models Spark Discussion (2026-08-10, 2 posts)

Developers are analyzing leaked architectures of two new AI models: a 400B parameter Mixture of Experts model with aggressive GQA, and a 30B parameter dense multimodal model. The community is highly anticipating the experimental potential of these architectures.

Episode 4 · llama.cpp Adds Day-0 Support for Meta's Muse Glimmer (2026-08-10, 2 posts)

The open-source inference framework llama.cpp announced Day-0 support for Meta's newly released Muse Glimmer model, with Meta officially providing GGUF quantized versions to facilitate immediate deployment.

Episode 5 · Meta's Open-Source Muse Glimmer 30B Impresses in On-Device Agent Tests (2026-08-11, 14 posts)

Meta's newly open-sourced Muse Glimmer 30B model has sparked intensive testing by developers. Optimized for agentic scenarios (multi-turn dialogue, tool calling), it demonstrates rigorous code handling and strong visual grounding, while achieving high-speed inference on consumer hardware and even in browsers, proving its potential as an on-device model.

Confirmed

  • Agentic and code capabilities: Blogger @altryne confirmed the model performs well in complex agentic scenarios with about 20 tools and large context. Developer @pbaylies' behavioral audit showed 0 hallucinated defects in 12 tests on working code, versus 10 for the baseline model.
  • Visual and UI grounding: Multiple developers verified multimodal accuracy on Mac Studio. Blogger @MaziyarPanahi tested the model successfully locating all 10 checkboxes in a synthetic medical form with accurate coordinates, and achieving an average IoU of 0.988 for element extraction in a scheduling UI test. Additionally, the model identified and located missing required fields in a referral form in 18.8 seconds, automatically suspending the workflow for human review.
  • Cross-hardware local inference performance: The model shows excellent compatibility and speed across devices. In Chrome on Mac Studio via WebGPU, generation speed reached 32 tokens/s; on M4 Max in browser, 25 tokens/s, which developers @xenovatech and @nicodotdev noted is comparable to native llama.cpp. On discrete GPUs, blogger @DanC403 achieved 20 t/s on an entry-level AMD RX 7600 XT (16GB) with speculative decoding; developer @curiousily reached 22 t/s on M5 Pro using 24GB memory; blogger @altryne also ran a GGUF quantized version on RTX 5090.

Why it matters

Muse Glimmer 30B's open-source release provides a viable on-device AI solution combining complex logic handling (multi-tool calling) and high-precision multimodal perception (accurate visual coordinate grounding). Its ability to run smoothly on consumer hardware and even in pure browser environments without high compute costs significantly lowers the barrier for local deployment in privacy-sensitive scenarios like healthcare and automated form processing.

Episode 6 · Muse Glimmer Tested: Reaches 92% of Qwen's Intelligence with 1/3 Compute (2026-08-11, 2 posts)

Recent tests reveal that the Muse Glimmer model achieves approximately 92% of Qwen 3.6 27B's intelligence while requiring only one-third of the compute power, demonstrating significant architectural efficiency despite slightly lower overall intelligence.

Episode 7 · Meta Releases Local Open-Source LLM as OpenAI Unveils Unrestricted Safety Model (2026-08-12, 2 posts)

Meta released a 30B parameter open-source model designed to run locally on consumer GPUs, while on the same day, OpenAI launched an unrestricted cybersecurity model for verified researchers. The moves highlight two distinct paths in AI development strategy.

Episode 8 · Single GH200 Hits 3323 tok/s in Muse Glimmer 30B Inference (2026-08-12, 3 posts)

A developer deployed the Muse Glimmer 30B model on a single NVIDIA GH200 GPU, achieving an extreme inference speed of 3,323 tokens/s using vLLM, BF16, and FlashAttention 3. The model's API is now available for testing.