Every figure, date, name and licence the video puts on screen, chased to a primary source.
Released 10 August 2026 by Meta, as an open-weight model. Meta AI Research announced it under the title “Introducing Muse Glimmer: An Open Agentic Model That Runs on Your Device”. Source: Meta AI Research, Introducing Muse Glimmer
30 billion parameters. Meta describes it as a “30-billion-parameter model optimized for always-on local agent workflows”. The model card on Hugging Face is more precise: approximately 29.6 billion parameters, a figure that includes the vision encoder. The video uses the 30B figure Meta itself publishes. Sources: Meta AI Research, Hugging Face model card, meta-models/Muse-Glimmer-30B
Apache 2.0. The weights are published under the Apache 2.0 licence, which permits commercial use, modification and redistribution. This matters for the open-weight section: the licence on the weights is permissive, and that is a separate question from whether the training data and the training pipeline were released. They were not. Sources: Hugging Face model card, gHacks, 11 Aug 2026
Quantized to roughly 4 bit, under 20 GB, from over 55 GB at full precision. Meta describes compressing the language model “to approximately 4-bit precision, shrinking the language model to under 20 GB” using K-Quant techniques, against “over 55 GB of memory” for a full-precision version. The model card lists two variants: K-Quant-Dynamic targeting a 32 GB card and K-Quant-17GB targeting a 24 GB card. Sources: Meta AI Research, Hugging Face model card
A 24 GB or 32 GB memory envelope, on a single consumer GPU. Meta states the model is designed to run on “a Mac or PC with a single consumer GPU” inside “a 24 GB or 32 GB envelope” once the KV cache, the perception encoder and the speculative decoding components are accounted for.
This is the figure the video is careful about. A 24 GB envelope is a high-end graphics card or a high-end Apple laptop, not an average consumer machine, so the shots state 24 GB rather than implying the model runs on any laptop. Source: Meta AI Research
Tested on an Nvidia RTX 5090, a MacBook M4 Max and a MacBook M5 Max. Those are the three machines Meta reports numbers for. Sources: Meta AI Research, Hugging Face model card
Speculative decoding speedups of 3.1x, 1.8x and 1.5x. Meta reports its DFlash speculative decoding achieving 3.1 times faster decode on the RTX 5090, 1.8 times on the M5 Max and 1.5 times on the M4 Max, against standard generation. Source: Meta AI Research
Distributed on Hugging Face.
The weights are a free download at huggingface.co/meta-models/Muse-Glimmer-30B.
Source: Hugging Face model card
Aimed at agentic work rather than at benchmark leadership. Meta lists local agents, function calling, local coding, LLM-as-a-judge evaluation, end-to-end agentic task completion, reliable tool use, multi-step reasoning and multimodal input. The model card also states plainly that Muse Glimmer “does not fall under the definition of ‘Frontier AI’” and is “generally less capable than Muse Spark”, which is the same point the video makes about local models not beating the frontier. Sources: Meta AI Research, Hugging Face model card
Meta’s closed flagship, released 5 August 2026, five days before Muse Glimmer. Announced alongside Muse Code, a terminal coding agent. Muse Spark 1.2 is served through the Meta Model API rather than downloaded. Sources: Meta AI Research, Introducing Muse Code and Muse Spark 1.2, MarkTechPost, 5 Aug 2026
Muse Glimmer is distilled from it. Meta describes training Muse Glimmer using logit distillation from Muse Spark’s outputs as the teacher model. This is why the small local model exists at all: it is the large closed model compressed, not an independent line of work. Source: Meta AI Research
Open weight means the parameters are published. It does not mean the training data or the training pipeline are. Meta published the Muse Glimmer weights under Apache 2.0. It did not publish the training corpus or a pipeline that would let anyone reproduce the model from scratch. That is the distinction the video draws, and it holds even though the licence on the weights is itself a permissive open source licence. Source: Hugging Face model card
Meta’s earlier open-weight releases became a major force in open AI. The video’s claim that Llama shaped tooling, startups and local AI support is a characterisation of a well documented history rather than a single sourced figure, and the video puts no number on it.