Gemma
Google's family of open-weight AI models you can download, run locally, and self-host
Meta's open-weight 30B agentic model that runs local, multimodal AI agents on a single GPU
Toolglade's take: Muse Glimmer is one of the most capable models you can genuinely run on your own hardware for free, and its agentic focus (tool use, failure recovery, multimodal input) makes it a standout for privacy-conscious local agents rather than just chat.
Muse Glimmer is Meta's 30-billion-parameter, Apache 2.0 open-weight model built for autonomous agents that run locally on consumer hardware. It combines multi-step reasoning, reliable tool calling, multimodal (text and image) input and failure recovery, and is distributed free on Hugging Face and popular local runtimes. Quantization and speculative decoding let it fit and run responsively on a single 24-32 GB GPU or a Mac.
Muse Glimmer is a 30-billion-parameter open-weight model from Meta Superintelligence Labs, released August 10, 2026 under a permissive Apache 2.0 license. Distilled from Meta's larger Muse Spark model, it is purpose-built for always-on, local agentic workflows: multi-step reasoning, reliable tool and function calling, multimodal (text and image) understanding, and failure recovery, all packaged into a single model that runs without cloud infrastructure or an internet connection. Through roughly 4-bit quantization and DFlash speculative decoding, Muse Glimmer compresses to under 20 GB and runs at practical speeds on a 24 GB or 32 GB consumer GPU or a Mac. Weights are available free on Hugging Face and through partners like Ollama, LM Studio, llama.cpp, MLX, ExecuTorch, vLLM and SGLang, plus hosted APIs from providers such as Together AI, Fireworks AI and OpenRouter for teams that prefer not to self-host.
Muse Glimmer is a 30-billion-parameter open-weight model from Meta Superintelligence Labs, launched August 10, 2026 under an Apache 2.0 license. It is distilled from Meta's Muse Spark and built for local, agentic workflows: tool use, multi-step reasoning, multimodal input and failure recovery. Quantization and speculative decoding let it run on a single 24-32 GB consumer GPU or a Mac. Weights are free on Hugging Face and popular local runtimes, with hosted APIs available from third-party providers.
Muse Glimmer is developed and released by Meta Platforms, specifically its Meta Superintelligence Labs research organization. Meta has a long track record of open AI research and releasing model weights to the developer community, and Muse Glimmer extends that tradition into local, agentic AI. The launch accompanied broader open-source messaging from the company, including plans to open the weights of its more advanced Muse Spark 1.2 model.
Muse Glimmer sits within Meta's wider Muse model family (Muse Spark, Muse Code, Muse Image) and is distributed through Meta's AI Developer Center and Hugging Face. As a Meta product it is not a standalone commercial offering with its own revenue model; the weights are given away free to drive adoption of Meta's open AI ecosystem.
Muse Glimmer is a dense causal transformer of roughly 29.6B parameters (including a ~1.8B ViT-G/14 perception encoder), with 52 layers, a 131,072+ token context window and a January 4, 2026 knowledge cutoff. It accepts interleaved text and image input and produces text output, and is trained on data from more than 100 languages. Its capabilities center on agents: end-to-end task completion, reliable schema-based tool and function calling, long-horizon multi-step reasoning, failure recovery, and compatibility with agentic scaffolds.
For local deployment, Meta quantizes the weights to roughly 4-bit precision (under 20 GB) and ships a DFlash speculative-decoding drafter that yields up to 3.1x faster generation on an RTX 5090. Released artifacts include full-precision BF16 weights, two 4-bit quant variants for 24/32 GB hardware, the drafter head and the perception encoder, all under Apache 2.0, with integrations across Hugging Face, Ollama, LM Studio, llama.cpp, MLX, ExecuTorch, vLLM and SGLang.
The primary audience is developers and technical teams building local or on-device AI agents, especially those with privacy, offline-operation or cost constraints that make cloud APIs unattractive. It also serves researchers who want a free, permissively licensed, fine-tunable model, and builders of coding agents, tool-use pipelines and multimodal agentic apps who need capable inference on a single consumer GPU or Mac.
Developers, ML engineers and researchers who download and run the model to build local agents, coding assistants and multimodal tools on their own hardware.
Because the weights are free under Apache 2.0, there is no traditional purchase; decision-makers are engineering leads or CTOs choosing to standardize on an open, self-hosted model to control cost and data privacy.
The open-source AI community, ML researchers, framework maintainers (Ollama, LM Studio, llama.cpp, vLLM) and tech commentators who benchmark and popularize open models.
A technically capable team or individual building agentic or on-device AI who values a free, permissively licensed, self-hostable model that runs locally on a single 24-32 GB GPU or high-end Mac, with strong tool-use and multimodal capabilities.
Not separately disclosed. Muse Glimmer is a product of Meta Platforms (via Meta Superintelligence Labs) and is funded as part of Meta's broader AI investment rather than as an independently financed entity.
Yes. Meta released the model weights free under the permissive Apache 2.0 license, which allows commercial and research use at no cost. You only pay for your own hardware if you self-host, or the per-token rates of a third-party hosted API provider if you use one.
The 4-bit quantized model compresses to under 20 GB and is designed to run within a 24 GB or 32 GB GPU memory envelope, so a single consumer GPU (such as an RTX 5090) or an Apple Silicon Mac (M4 Max / M5 Max) is enough. Full-precision BF16 weights require roughly 64 GB of VRAM.
Muse Glimmer is Meta's newer agent-focused open model (Meta positions it in the lineage of its open releases rather than the Llama name). In Meta's own published benchmarks it is compared against similarly sized open models Gemma4-31B and Qwen3.6-27B, where it leads on many agentic and coding tasks while trading places with Qwen on some multimodal and reasoning benchmarks. It is optimized specifically for local agentic use.
Muse Glimmer is released as open weights under the Apache 2.0 license, a permissive license with no additional usage restrictions on commercial deployment. All released artifacts, including the full-precision weights, 4-bit quants, DFlash drafter and perception encoder, are covered by Apache 2.0.
Download the weights from the official Hugging Face repo (meta-models/Muse-Glimmer-30B), pick a 4-bit quant for consumer GPUs, then load it in a local app like Ollama, LM Studio or llama.cpp, or serve it with vLLM/SGLang for an OpenAI-compatible API. Optimized builds for MLX and ExecuTorch support Apple Silicon and edge devices.
Side-by-side pages for pricing, features, and best-fit use cases.
This week in AI: Meta's open-weight Muse Glimmer, an escalating model price war, Perplexity Comet goes free, OpenAI retires Atlas, and a wave of inference funding.
Local AI went mainstream in 2026. A developer's guide to the best open-weight models to run on your own GPU or Mac: Muse Glimmer, Gemma, Qwen, DeepSeek, and GLM.
Google's family of open-weight AI models you can download, run locally, and self-host
Open-weight frontier LLM family from Z.ai (Zhipu AI), tuned for coding and agents.
Alibaba's free AI assistant, backed by the open-weight Qwen model family
Free AI chat assistant from Moonshot AI, built for long-context reasoning and open Kimi K2 models