Meta Muse Glimmer-30B
One-line: Meta’s return to open weights — a 30B dense agentic model (VL input) purpose-built to run on consumer hardware (K-Quant ~17GB fits 24GB VRAM), with published SWE-bench Verified 76.0. 🔬 VL model with real coding/agentic evals → included per tracker rule.
ℹ️ Released Aug 9-10, 2026 by Meta Superintelligence Lab under org
meta-models(notmeta-llama). Confirmed real via research.meta.ai blog “Introducing Muse Glimmer: An Open Agentic Model That Runs on Your…”, NVIDIA developer blog, HF blog (Aug 10), marktechpost, r/LocalLLaMA. HF download counter shows 0 (counter lag — 1,260 likes + unsloth/bartowski/RedHatAI quant repos in 24h are inconsistent with 0 downloads); weights verified present (2× safetensors shards, ~60GB fp16). ⚠ All benchmark numbers below are lab-reported (Meta’s card / HF blog, backed by research.meta.ai methodology report). No independent re-run yet.
Coding benchmarks
| Benchmark | Score | Source |
|---|---|---|
| HumanEval | — | not published |
| HumanEval+ | — | not published |
| MBPP | — | not published |
| LiveCodeBench | — | not published |
| Aider polyglot | — | not published |
| SWE-bench Verified | 76.0 | lab-reported |
| SWE-bench Pro | 51.2 | lab-reported |
| TerminalBench 2.1 (with terminus2) | 51.7 | lab-reported |
| SciCode | 43.6 | lab-reported |
Agentic / tool-use benchmarks
| Benchmark | Score | Source |
|---|---|---|
| BFCL | — | not published |
| τ-Bench | — | not published (uses τ³-Bench instead) |
| τ³-Banking | 23.5 | lab-reported |
| MCP Atlas (Public) | 75.5 | lab-reported |
| DeepSearch QA | 74.6 | lab-reported |
| GAIA2 | 43.3 | lab-reported |
| OSWorld-Verified | 65.9 | lab-reported |
| WildClawBench | 47.6 | lab-reported |
Reasoning (context only)
| Benchmark | Score | Source |
|---|---|---|
| MMLU / MMLU-Pro | — | not published |
| GPQA Diamond | 83.5 | lab-reported |
| AIME 2026 | 94.7 | lab-reported |
| HLE Text | 22.0 | lab-reported |
| IFBench | 77.0 | lab-reported |
| MMMU Pro | 74 | lab-reported |
Efficiency
- Active params: 30B dense (29.6B incl. 1.8B ViT-G/14 perception encoder)
- Context length: 131,072+ (120K+ per NVIDIA blog; max visual tokens 4,096/image)
- VRAM (fp16 rough): ~60 GB. The release is explicitly quantization-first: K-Quant-Dynamic → 32GB VRAM, K-Quant-17GB → 24GB VRAM (validated ≤1.0% degradation across 15 benchmarks). DFlash speculative decoder (16-token blocks): RTX 5090 74.9 → 233.4 tok/s (3.1×); M4 Max 23.7 → 37.8; M5 Max 26.6 → 50.2 (ExecuTorch).
llama serve -hf meta-models/Muse-Glimmer-30B-GGUFand transformersAutoModelForMultimodalLMboth supported.
What makes it notable
This is the first open-weight Meta model explicitly designed for always-on local agent workflows: multi-step reasoning with controllable effort (low/med/high/xhigh reasoning strength), reliable tool use, failure recovery after bad tool calls, and interleaved text+image input (screenshots/charts/documents) in one 24GB-VRAM package. On Meta’s numbers it beats Gemma4-31B on every agentic row (MCP Atlas 75.5 vs 54.2, DeepSearch QA 74.6 vs 61.7, SWE-bench Verified 76.0 vs 66.6, τ³-Banking 23.5 vs 15.1) and roughly matches Qwen3.6-27B while being natively multimodal. The card explicitly lists OpenClaw and Hermes Agent as supported scaffolds — for this user (local runner, pipes models into agents) this is the most directly relevant release of the week. Distribution via GGUF (meta-models + unsloth + bartowski) and ExecuTorch-PTE for Apple Silicon. Apache-2.0 — no use restrictions. First open-weight Meta release since the Llama-4 series.