AIBRIEF

Meta Returns to Open Source With Muse Glimmer, a 30B Apache 2.0 Agent Model

N
NathanTechnology Editor · Technical Lead
Published · Updated
According to VentureBeat, Meta released Muse Glimmer on August 10, 2026 — a 30-billion-parameter, Apache 2.0-licensed open-weight model built to run AI agents on consumer hardware. TechCrunch reports it is an open version of Meta's closed Muse Spark model, framed by CEO Mark Zuckerberg as evidence of his "personal superintelligence" vision.

Why Meta Opened Glimmer's Weights

Meta released Muse Glimmer on August 10, 2026, described by VentureBeat as "a 30-billion-parameter open-weight model designed to run autonomous AI agents directly on consumer hardware" (E1). The release carries the Apache 2.0 license — VentureBeat calls it "the company's first fully open release since it succeeded its open-weight Llama family in April with the proprietary Muse Spark" (E2).

Meta co-founder and CEO Mark Zuckerberg announced the move on X: "Today we're also opening the weights for Muse Glimmer, a great 30B parameter dense model that can run locally," adding that Meta would "soon" release weights for Muse Spark 1.2 and calling himself "a strong supporter of open source" (E4). TechCrunch frames the release as "the clearest picture yet of what CEO Mark Zuckerberg's vision of 'personal superintelligence' could look like in practice" (E16), and reports that in a letter published the same Monday, Zuckerberg argued that distributing superintelligence widely "has the potential to begin a new era of personal empowerment where individuals can use this powerful new capability to reach their full potential, pursue their interests, and improve their lives and the world more than ever before" (E19).

Inside Glimmer's 30B Architecture and Speed Optimizations

Per VentureBeat's review of the Hugging Face model card, Glimmer is "a dense causal transformer with approximately 29.6 billion total parameters across 52 layers, including a dedicated ~1.8B-parameter ViT-G/14 perception encoder," with a stated context length of 131,072 tokens or more and a knowledge cutoff of January 4, 2026 (E6). The weights are live on Hugging Face now, with VentureBeat reporting that support is rolling out this week through Ollama, LM Studio, vLLM, SGLang, Together AI, Fireworks AI and OpenRouter (E3). TechCrunch adds that Glimmer is designed to run multi-step agent tasks — calling tools, writing and debugging code, working with files and screenshots — locally on a Mac or PC with a single consumer GPU, supports text and images, and was trained across more than 100 languages (E18).

Memory footprint is a central design constraint. VentureBeat reports that at full precision, the 30B model requires more than 55GB of memory, prompting Meta to develop roughly 4-bit quantized versions that shrink the language-model weights to under 20GB (E7):

ConfigurationTarget hardwareAccuracy degradation (15 benchmarks)
Full precision— (>55GB memory)
K-Quant-Dynamic32GB hardware0.2%
K-Quant-17GB24GB hardware1%

Meta itself cautions that "those figures are Meta's own measurements, not independent evaluations" (E8). Meta also reports speed gains from its DFlash speculative-decoding technique:

HardwareBaseline (tok/s)With DFlash (tok/s)Speedup
Nvidia RTX 509074.9233.43.1x
Apple M5 Max26.650.21.8x
Apple M4 Max23.737.81.5x

(E9)

How Glimmer Fits Into Meta's Muse Lineup

Alexandr Wang, Meta's chief AI officer, wrote on X that "just like much larger models, muse glimmer can operate as a fully capable agent via planning, tool calls, checking its own results, and failure recovery," and that the model "can run on 24GB of VRAM without losing agentic reliability" (E5). TechCrunch reports that Glimmer "is essentially an open version of Meta's most powerful closed model, Muse Spark, which the company debuted in April" (E17), and that while "Muse Spark, its more powerful model, remains closed-weight," the smaller Glimmer "can be downloaded, fine-tuned, and run on a user's hardware" (E20).

According to VentureBeat, Meta evaluated Glimmer under its Advanced AI Scaling Framework and determined it does not meet the framework's definition of "Frontier AI" because it is "generally less capable than Muse Spark"; Meta's Preparedness Team assessed the model at Moderate or lower risk across chemical/biological, cyber and loss-of-control categories (E15).

How Glimmer Stacks Up Against Rival Open-Weight Models

VentureBeat reports that in Meta's own three-way comparison, Glimmer leads Gemma4-31B and Qwen3.6-27B on several agentic tests, but Qwen leads on others:

BenchmarkGlimmerGemma4-31BQwen3.6-27B
MCP Atlas75.5
DeepSearch QA74.6
τ³-Banking23.5
WildClawBench47.6
GAIA243.3
SWE-Bench Pro51.236.950.2
SWE-Bench Verified76.077.2
OSWorld-Verified65.975.6
TerminalBench 2.151.760.7
GDPval-AA9531141

Glimmer tops MCP Atlas, DeepSearch QA, τ³-Banking, WildClawBench, GAIA2 and SWE-Bench Pro (E10), while Qwen "leads Meta's own comparison on OSWorld-Verified (75.6 vs. Glimmer's 65.9), TerminalBench 2.1 (60.7 vs. 51.7), SkillsBench, GDPval-AA (1141 vs. 953) and most of the multimodal benchmarks," with Glimmer's SWE-Bench Verified score of 76.0 landing "just below Qwen's 77.2" (E11).

Privacy and Prompt-Injection Risk, Quantified

On the CI Memories privacy benchmark — where lower violation rates are better — VentureBeat reports Glimmer scores 26.4, against Gemma's 12.1 and Qwen's 53.4 (E13). On the Siren AgentDojo prompt-injection test, Glimmer shows a 28.4% attack-success rate versus 25.6% for Gemma and 40.3% for Qwen, while posting the highest utility score of the three at 94.2 (E14).

BenchmarkGlimmerGemmaQwen
CI Memories violation rate (lower = better)26.412.153.4
Siren AgentDojo attack-success rate28.4%25.6%40.3%
Siren AgentDojo utility score94.2

The Open-Weight Market Glimmer Is Entering

VentureBeat cites OpenRouter data showing that by May 2026, Chinese open-weight models accounted for roughly 61% of all tokens consumed on OpenRouter, with four of the five most-used models coming from Chinese labs — while Meta's Llama, the prior open-weight leader, fell off the rankings entirely (E12).

What This Means

The evidence lays out a reversal with conditions attached. Meta closed its most capable model, Muse Spark, in April (E17, E20) and only four months later opened a smaller sibling under Apache 2.0 (E1, E2) — timed to a moment when, per OpenRouter data cited by VentureBeat, Llama had already dropped out of the top open-weight rankings and Chinese labs commanded roughly 61% of token consumption (E12). Meta's own benchmark disclosures show a split picture rather than outright superiority: Glimmer leads Qwen3.6-27B and Gemma4-31B on several agentic tasks (E10) but trails Qwen on OSWorld-Verified, TerminalBench 2.1 and GDPval-AA (E11), and Meta explicitly labels its quantization accuracy figures as self-reported rather than independently verified (E8). The same pattern holds on safety: Glimmer's CI Memories violation rate and AgentDojo attack-success rate both sit between Gemma's lower figures and Qwen's higher ones (E13, E14), positioning Glimmer as a middle-of-the-pack result on the risk metrics Meta itself published alongside its Advanced AI Scaling Framework assessment (E15).

📊 Evidence

FAQ

What license does Muse Glimmer use?

Apache 2.0, according to VentureBeat — described as "the permissive, industry-standard Apache 2.0 open source license," Meta's first fully open release since Llama.

How much memory does Glimmer need to run?

Meta's chief AI officer Alexandr Wang said Glimmer can run on 24GB of VRAM without losing agentic reliability; VentureBeat reports the full-precision model needs over 55GB, while Meta's ~4-bit quantized builds shrink the language-model weights to under 20GB.

Where can developers get Glimmer today?

Weights are live on Hugging Face now, with VentureBeat reporting support rolling out this week via Ollama, LM Studio, vLLM, SGLang, Together AI, Fireworks AI and OpenRouter.

Does Glimmer beat Qwen3.6-27B?

Results are mixed in Meta's own testing, per VentureBeat: Glimmer leads on MCP Atlas, DeepSearch QA, τ³-Banking, WildClawBench, GAIA2 and SWE-Bench Pro, but Qwen leads on OSWorld-Verified, TerminalBench 2.1, SkillsBench, GDPval-AA and most multimodal benchmarks.

📎 Sources

  1. venturebeat.com
  2. techcrunch.com
N
NathanTechnology Editor · Technical Lead

Related

BRIEF

Anthropic Makes Claude Code's Auto Mode the Default Starting August 14

According to TechCrunch and ITHome, Anthropic will set auto mode as the default permission setting for Claude Code Pro, Max, and Team accounts from August 14, 2026, while Enterprise, API, and cloud-platform users must opt in manually. Anthropic's internal study of 1,053 paid testers found auto mode blocked 89% of harmful actions versus 13.6% for manual human review, and separately reported that users had been approving 97% of permission prompts before this change.

Nathan ·
BRIEF

NVIDIA Partners With Six Wall Street Firms on $500 Billion AI Financing Platform

NVIDIA (輝達) has signed memorandums of understanding with six Wall Street firms — Apollo, BlackRock (貝萊德), Blackstone, Brookfield, Goldman Sachs (高盛), and KKR — to launch a computing financing platform targeting over $500 billion in third-party funding for AI infrastructure, according to CNA (中央社) and Liberty Times (自由時報) reports. The move follows Moody's warning that unprecedented capital spending is squeezing tech firms' free cash flow, per TechNews (科技新報). NVIDIA has not disclosed financing terms, individual commitment amounts, or a deployment timeline for the $500 billion target.

林紀旭 James Lin ·
BRIEF

SK Hynix Board Approves 54 Trillion Won for New Yongin and Cheongju Wafer Fabs

SK Hynix's board approved a combined 54 trillion won (about $38 billion) investment in two new fabs — Yongin Y2 (35.2 trillion won) and Cheongju M17 (19.1 trillion won) — according to reports from TechNews and Inside.com.tw dated August 7, 2026. The move follows Omdia's forecast of 19% average annual DRAM and NAND demand growth through 2030.

林紀旭 James Lin ·