Muse Glimmer Returns Meta to Open Weights — AA 35, Not Sol-Class
Meta released Muse Glimmer, a 30B Apache 2.0 open-weight model for always-on local agents on a single consumer GPU, alongside Zuckerberg’s open-weight policy essay. Artificial Analysis scores it 35 on Intelligence Index — 21 points above Llama 4 Maverick — local agent throughput, not cloud frontier Elo.
Open-weight competition is no longer a China-only story. Apache 2.0 at 30B for offline agents answers cost, cyber-refusal, and sovereignty concerns that closed APIs create — while Zuckerberg’s same-day essay frames open weights as U.S. industrial policy.
Meta shipped Muse Glimmer, a 30-billion-parameter dense model from Meta Superintelligence Labs, with weights under Apache 2.0 on Hugging Face. Built for always-on local agent workflows — function calling, local coding, multi-step tool use, LLM-as-a-judge — on a Mac or PC with a single consumer GPU (quantized under 20 GB with headroom for KV cache and perception). Same day, CEO Mark Zuckerberg published a **14-page** essay (“The Future is for Everyone”) urging U.S. policy to ease open-weight friction versus Chinese labs and criticizing extreme concentration of AI power — without naming OpenAI and Anthropic.
What shipped
| Item | Detail (Meta research blog / HF / AA) |
|---|---|
| Model | Muse Glimmer — 30B dense; multimodal text + images via perception encoder |
| License | Apache 2.0 (Meta’s most permissive open release yet vs prior Llama License) |
| Context | 128K token context (plus extension) |
| Target | Local agents: long-horizon tasks, precise tool schemas, failure recovery / retry |
| Memory path | Full precision >55 GB; ~4-bit quant brings LM under ~20 GB for 24–32 GB consumer envelopes |
| Speed stack | Speculative decoding with DFlash-style drafter; validated on M4/M5 Max Macs and RTX 5090 |
| Training | Logit distillation from Muse Spark; mid-train agent traces; SFT + on-policy distillation + RL |
| Access | meta-models/Muse-Glimmer-30B on Hugging Face; Ollama / LM Studio / llama.cpp / MLX / ExecuTorch integrations landing |
| Safety | Evaluated under Meta’s Advanced AI Scaling Framework for open-weight release |
Zuckerberg’s messaging also previewed opening weights for Muse Spark 1.2 and “even bigger models coming soon,” plus a $1 billion community fund tied to data-center build-out — distinct product lines from the Glimmer ship itself.
Claims vs checks
Lab-reported: Meta positions Glimmer as strong vs Gemma 4 31B and Qwen3.6-27B on agentic suites (DeepSearch QA, MCP-Atlas, τ-Bench, SWE-Bench) with methodology notes in the launch report.
Independent (near launch):
- Artificial Analysis (Aug 10): Muse Glimmer scores 35 on the Intelligence Index — 21 points above Llama 4 Maverick (14); ~5 points above Gemma 4 31B (Reasoning) in AA’s size-class comparison. Openness Index ~44.
- LMArena / Arena: live boards may lag a same-day open drop; do not invent Elo.
- Not a frontier closed-model challenger on AA’s top band (Opus / Sol class); the thesis is local agent throughput + open weights.
Distinct from Muse Code / Spark 1.2 (Aug 5 closed coding stack) and the Aug 5 cyber-eval breach. Glimmer is Meta’s first open-weights MSL release since Llama 4.
Limits
- Spark 1.2 weights and “bigger models” are previewed, not shipped here.
- Arena Elo not claimed for a same-day drop.
- AA 35 is size-class open strength, not Sol/Opus peer territory.
Sources
- Meta AI Research: Introducing Muse Glimmer (August 10, 2026)
- Hugging Face: meta-models/Muse-Glimmer-30B
- Artificial Analysis: Muse Glimmer benchmarks (August 10, 2026)
- Reuters: Meta launches new AI model as Zuckerberg champions open-weight push (August 10, 2026)
- CNBC: Meta Muse Glimmer open-weight AI (August 10, 2026)
- Times of AI:
meta-muse-code-spark-1-2(August 5)