Meta's Muse Glimmer: what a 30B open-weight agent model changes
· TexaHub · LLMs & Foundation Models, Artificial Intelligence, Model Comparisons
Meta's first open-weight release since Llama 4 targets local agents on one consumer GPU; its benchmarks, hardware claims and safety numbers deserve a closer look.
What Meta released
On 10 August Meta published the weights of Muse Glimmer, a model of roughly 30 billion parameters that it describes as built for local agents, under the Apache 2.0 licence [2]. Meta calls Apache 2.0 "the most permissive license we've used for an open model" [1]. The weights are distributed through Hugging Face [1], and the model card lists it as intended for "commercial and research use" [3].
The release marks a change of direction. Meta's previous major open models were Llama 4 Scout and Maverick, released in April 2025 [12]. In April 2026 the company introduced Muse Spark, the first model in its new Muse series, and offered it through the Meta AI app and a private API preview for selected partners rather than as downloadable weights, while saying it hoped to open-source future versions [11]. Muse Spark remains closed-weight, and TechCrunch described Glimmer as essentially an open version of that model sized for consumer hardware [4].
Mark Zuckerberg framed the release in terms of openness: "Today we're also opening the weights for Muse Glimmer, a great 30B parameter dense model that can run locally," he said [6]. Constellation Research reported that an open-weight version of Muse Spark 1.2 is also planned [8], though Meta has not given a date for it in the material we reviewed.
The launch was paired with a signed essay by Zuckerberg arguing that distributing superintelligence widely "has the potential to begin a new era of personal empowerment" [4]. Thurrott reported that the announcement came alongside plans to invest $145 billion in data centres this year and a $1 billion fund aimed at local opposition to infrastructure projects [10].
Inside the model
The model card describes a dense causal transformer of about 29.6 billion parameters in total, including a vision encoder of about 1.8 billion parameters, with 52 layers and a hidden dimension of 6,656 [3]. Hugging Face's launch post describes the same system as a 2B vision encoder paired with a 28B text decoder [2], so the published component sizes differ slightly depending on the source. The model accepts text and images as input and produces text [3].
Its attention pattern repeats three local layers followed by one global layer, with a 2,048-token sliding window [3]. Hugging Face describes the attention as gated grouped-query attention with a 16:1 ratio of query to key-value heads [2]. The model card gives a context length of 131,072 or more tokens, a vocabulary of 202,048 tokens and a knowledge cutoff of 4 January 2026 [3], while Meta's developer blog refers to a 128K-token default context window [1].
On training, Constellation Research reported that Glimmer was distilled from Muse Spark's outputs, followed by training on longer-context and agent-heavy data using policy distillation and reinforcement learning [8]. Engadget reported that the model is based on Spark 1.2 and was trained on data from more than 100 languages [7]. VentureBeat noted that Meta has not released the training data or training code [6], which places Glimmer in the open-weight rather than fully open-source category, whatever the licence.
Meta's developer blog emphasises agent behaviour over raw chat quality, saying that "tool calling holds up over long sequences" and that the model "recovers instead of halting" when a tool call fails [1]. These are design goals stated by the company; independent evaluations over long, real-world workflows were not yet available in the sources reviewed for this report.
Benchmarks and their limits
Meta compared Glimmer chiefly with two open models of similar size, Gemma4-31B and Qwen3.6-27B [8]. On MCP Atlas, Glimmer scored 75.5 against 54.2 for Gemma4-31B and 62.5 for Qwen3.6-27B [9]. On SWE-Bench Pro the scores were 51.2, 36.9 and 50.2 respectively [6]. MarkTechPost also listed scores of 74.6 on DeepSearch QA, 43.3 on Gaia2, 94.7 on AIME 2026 and 77.0 on IFBench [9].
The comparison is not one-sided. On SWE-Bench Verified, Glimmer scored 76.0 against 77.2 for Qwen3.6-27B; on OSWorld-Verified it scored 65.9 against 75.6; and on TerminalBench 2.1 it scored 51.7 against 60.7 [6]. In other words, Meta's published table shows Qwen ahead on computer-use and terminal tasks, while Glimmer leads on MCP Atlas and SWE-Bench Pro.
All of these figures are vendor-reported, and VentureBeat cautioned that benchmark advantages still need to be validated in real enterprise conditions [6]. The choice of comparators also matters: a 30B-class model is being measured against peers of similar size, not against Meta's own closed Muse Spark or other frontier systems, so the results say more about Glimmer's position among locally runnable models than about the frontier.
Running it locally
The central practical claim is that Glimmer runs on one consumer GPU. Meta says its quantised version is under 20 GB, fitting on GPUs with 24 to 32 GB of memory, and that the model supports CPU, mixed CPU and GPU, and full GPU execution [1]. According to VentureBeat, full BF16 precision needs 55 GB or more of memory, and Meta reports accuracy degradation from 4-bit quantisation of 0.2% for the 32 GB variant and 1% for the 24 GB variant [6].
Meta also ships a small drafter model for speculative decoding [1]. VentureBeat reported generation speeds with this DFlash method rising from 74.9 to 233.4 tokens per second on an RTX 5090, from 26.6 to 50.2 on an M5 Max and from 23.7 to 37.8 on an M4 Max [6]. Meta's blog recommends drafting five tokens per step as a starting point, eight to ten for code and three to four for open-ended generation [1].
Tooling support is broad but was still rolling out at launch. Meta's developer blog lists vLLM, llama.cpp, ExecuTorch, Ollama, LM Studio, Unsloth and SGLang as deployment options [1], while Thurrott reported on launch day that Ollama and LM Studio availability was "coming soon" [10]. For fine-tuning, Hugging Face lists LoRA on a single 80 GB H100 and full supervised fine-tuning on eight 80 GB GPUs [2].
The hardware story is the most consequential part of the release for enterprises. A model that fits on a single workstation-class GPU can be run on premises, offline and without per-token fees, which addresses data-residency and cost concerns that keep some organisations away from hosted APIs. The trade-off is that the deploying organisation takes on responsibility for guardrails, monitoring and updates.
Safety and caveats
Meta published safety results alongside capability scores. On the Siren AgentDojo measure, Glimmer's attack success rate was 28.4%, against 25.6% for Gemma4-31B and 40.3% for Qwen3.6-27B [6]. On the CI Memories privacy test, Glimmer recorded 26.4 violations against 12.1 for Gemma and 53.4 for Qwen [6]. On both measures Glimmer sits between its two comparators rather than leading.
VentureBeat reported that Meta places the model's risk at moderate or lower and does not classify it as frontier AI, and that the company recommends deploying it with guardrails and human-in-the-loop confirmation for irreversible actions [6]. The model card warns that it "may produce inaccurate, biased, or objectionable responses" and "may still make errors in multi-step reasoning", and restricts use to people aged 18 and over [3].
For an agent model, the attack success rate is the figure that deserves most attention. A model that executes tool calls on a user's own machine, with access to files and applications, and that is compromised in more than a quarter of attack attempts in that test, needs external controls before it is trusted with sensitive actions. That is consistent with Meta's own deployment advice.
The strategic reading
TechCrunch characterised the release as Meta drawing a sharper line between the models it will let people own and the more capable ones it keeps under its control [4]. Meta Research argued that "smaller models, when trained effectively, can approach frontier-level performance on targeted tasks" [8], and Zuckerberg said that "rather than centralizing superintelligence, we should distribute it widely and give every person the ability to direct it" [7].
Not all reaction was favourable. TechCrunch's AI editor Russell Brandom argued that the accompanying manifesto failed to engage with documented harms, writing that "the public does not trust tech executives to make sure new technologies like this have a positive impact on society" [5]. That criticism is aimed at the essay rather than the model, but it shapes how a consumer-facing agent strategy is likely to be received.
Our assessment is that Glimmer is best read as a distribution play. Meta keeps its most capable system closed, while seeding developers with a permissively licensed model built for the local agent workloads it expects to matter. Whether it gains lasting adoption will depend on independent tests of the agent claims, on how quickly the promised tooling arrives and on whether the planned open Muse Spark 1.2 follows. None of those questions can be answered ten days after launch.
Sources
[1] Muse Glimmer: an open-weight model built for local agents — Meta, 12 Aug 2026
[2] Meta is back with Muse Glimmer: local, agentic, multimodal, and open source — Hugging Face, 10 Aug 2026
[3] Muse-Glimmer-30B model card — Hugging Face, undated
[4] Meta's new Glimmer AI model offers a hint at Zuckerberg's personal intelligence vision — TechCrunch, 10 Aug 2026
[5] Mark Zuckerberg's AI manifesto is exactly why people don't like AI — TechCrunch, 10 Aug 2026
[6] Meta returns to open source with Muse Glimmer, an Apache 2.0 licensed 30B parameter AI model optimized for agents — available now — VentureBeat, 10 Aug 2026
[7] Meta's 'open source' Muse Glimmer model can run on a single computer — Engadget, 10 Aug 2026
[8] Meta releases open weight Muse Glimmer model with open Muse Spark 1.2 on tap — Constellation Research, 10 Aug 2026
[9] Meta AI Releases Muse Glimmer: A 30B Open-Weights Agentic Model That Runs on One Consumer GPU — MarkTechPost, 10 Aug 2026
[10] Meta Releases Open Source, Open Weight Version of its Most Powerful AI Model — Thurrott.com, 10 Aug 2026
[11] Introducing Muse Spark: Meta's Most Powerful Model Yet — Meta, 8 Apr 2026
[12] The Llama 4 herd: The beginning of a new era of natively multimodal AI innovation — Meta, 5 Apr 2025