Meta Muse Glimmer 30B: Meta Is Back, And They Shipped The Local Agentic Model That Llama 4 Should Have Been
Meta Muse Glimmer 30B: Meta Is Back, And They Shipped The Local Agentic Model That Llama 4 Should Have Been
Hey guys, Mr. Technology here.
As of yesterday — August 10, 2026 — Meta Superintelligence Lab quietly shipped meta-models/Muse-Glimmer-30B on Hugging Face under Apache 2.0. Day-zero support landed simultaneously in transformers, llama.cpp, vLLM, and HF Inference Endpoints. The dense 30B multimodal model is distilled from the larger Muse and was designed from day one for local agentic use. It posts benchmark numbers that put it above Gemma 4 31B and Qwen 3.6 27B on the agentic tasks that matter for production.
This is not a model release. This is Meta reclaiming the open-source throne at the size class that actually runs on your laptop.
Why 30B Is The New Sweet Spot
Llama 3.1 405B was impressive and useless to anyone without a multi-GPU rig. The "open-source frontier" drifted to a size class developers could not actually run. The Llama 4 cycle continued that drift — bigger, more expensive, less aligned — and the community quietly routed around Meta to DeepSeek, Qwen, and Mistral.
Muse Glimmer makes a different bet: dense 30B, multimodal, Apache 2.0, designed to run on one box. At 4-bit that is ~16-20GB. That fits a single RTX 4090, a Mac Studio M2 Ultra at bf16, a Strix Halo laptop at Q4. Open source does not matter if you cannot run it. Meta just put a credible open-source agentic model back into that window.
The Architecture
HF's blog gives you the marketing. Here is the read that matters for someone shipping this.
Text decoder (28B). Hybrid attention alternating (SWA, SWA, SWA, Full) for 52 layers — three sliding-window layers (window 2,048, RoPE with θ = 500,000) then one full-attention layer with NoPE, repeated 13 times. SWA layers do 95% of the work cheaply; full-attention layers preserve global information every fourth layer. Beam 128K (long-context retrieval) comes in at 65.1 — ahead of Gemma 4 31B's 58.2 and competitive with Qwen 3.6 27B's 63.0.
Gated GQA — 16 KV heads shared across 32 query heads. A 16x reduction in KV cache memory vs vanilla MHA at this parameter count. For agentic workloads — long context windows of tool outputs, MCP responses, conversation history — KV cache is the bottleneck.
Q-K normalization with extra query scaling. Every query and key head gets RMS-normalized before attention, then queries are scaled to set the target logit scale. Keeps attention logits stable. Fewer "model got lost in the middle" failures at long context. Real win for any agent holding a 100K-token conversation.
Perception Encoder (2B). A 2B ViT-style encoder based on Meta's Perception Encoder, 50 layers, GELU MLPs, 2D RoPE. Handles images AND video through the same encoder — 2 fps, capped at 96 frames. Charxiv Reasoning at 78.8 leads the 31B-and-below class. Real multimodal.
DFlash speculative decoding drafter. Optional — disable to save memory, enable for ~2x faster generation at some VRAM cost. Particularly suited to structured output like coding and tool calls. On an RTX 5090, 74.9 tok/s baseline jumps to 233.4 tok/s with DFlash — a 3.1x speedup.
The Benchmarks (What Actually Matters)
Marketing wants MMLU. Practitioners care about the agentic numbers.
| Benchmark | Glimmer 30B | Gemma 4 31B | Qwen 3.6 27B |
|---|---|---|---|
| MCP Atlas (tool use) | 75.5 | 54.2 | 62.5 |
| DeepSearch QA | 74.6 | 61.7 | 71.1 |
| WildClawBench | 47.6 | 37.6 | 43.2 |
| GAIA2 | 43.3 | 36.4 | 40.0 |
| SWE-Bench Pro | 51.2 | 36.9 | 50.2 |
| SWE-Bench Verified | 76.0 | 66.6 | 77.2 |
| OSWorld-Verified | 65.9 | 58.5 | 75.6 |
| τ³-Banking | 23.5 | 15.1 | 16.7 |
| Charxiv Reasoning | 78.8 | 77.7 | 78.4 |
Muse Glimmer dominates the agentic benchmarks that matter for production. MCP Atlas — the most representative tool-use benchmark the industry has — leads at 75.5, +13 over Gemma and +13 over Qwen. SWE-Bench Pro (the harder subset) leads at 51.2. DeepSearch QA, WildClawBench, GAIA2, τ³-Banking — all leads for Glimmer.
Where it loses: OSWorld-Verified (Qwen wins on computer-use at 75.6) and SWE-Bench Verified (Qwen 77.2 vs Glimmer 76.0) — coding-heavy benchmarks where Qwen has been the strongest open-source model for the past year.
Summary: Muse Glimmer is the best open-source agentic model at 30B and below for tool use, multimodal reasoning, and long-horizon orchestration. Qwen 3.6 27B is still the strongest pure coding model. If your workload is research, tool orchestration, or multimodal — Glimmer wins.
Quantization And Local Deployment
Meta shipped calibrated GGUFs directly. K-Quant-Dynamic targets 32GB VRAM at 0.2% degradation. K-Quant-17GB targets 24GB VRAM (a single 4090) at 1.0% degradation. Both variants are within rounding error of bf16 on 15 common benchmarks.
The drafter quant is included in the release, so you can run main model + drafter + KV cache + perception encoder all inside a 24GB envelope. The math: 30B agentic, multimodal, with speculative decoding, on a single consumer GPU. Unsloth is also shipping optimized quants.
The Apache 2.0 Bet
Llama 3 / Llama 4 shipped under the Llama Community License — Meta-specific clauses, commercial-use restrictions above 700M MAUs, a six-week legal review for any global product. Muse Glimmer ships under Apache 2.0. No Meta-specific clauses. No user thresholds. No restrictions on derivative works. Same license DeepSeek V3, Qwen 3.x, Mistral, and Gemma use.
For one operator building production systems: Apache 2.0 is the license your legal team approves in one afternoon. Meta is competing on what matters — model quality and ecosystem — not on license friction.
What You Ship This Week
1. Pull the model and benchmark your actual workload.
pip install --upgrade transformers accelerate huggingface-cli download meta-models/Muse-Glimmer-30B-GGUF
Run it against your top three agentic use cases. Serve it with llama serve -hf meta-models/Muse-Glimmer-30B-GGUF for an OpenAI-compatible endpoint.
2. Test the DFlash drafter on coding workloads. ~2x speedup at the cost of some memory. Day-zero support exposes both the main model and the drafter through AutoModelForMultimodalLM.
3. Replace your Llama 3.x fallback with Glimmer. Better on agentic, multimodal-capable, same memory footprint at Q4. If you were holding out for Qwen 3.6, this is your "open source LLM got better this week" notification.
4. Re-price the multimodal tier. Any workload currently routing image Q&A to a 200B closed model can be retargeted to a 30B open model running locally. The cost structure flips from per-token to amortized-to-zero.
The Take
Meta just made the open-source bet a lot smarter. Muse Glimmer 30B is the first open-source LLM since Llama 3.1 405B competitive with closed-source frontier on the benchmarks that matter for production agentic systems. It runs on one box. Multimodal. Apache 2.0. Day-zero ecosystem support.
The bigger play: Meta is not trying to beat Anthropic and OpenAI at the closed-source frontier. They are trying to make that frontier irrelevant for the workloads that can run on local hardware. Tool use, document analysis, multimodal Q&A, coding assistance — these are the workloads where a 30B open-source model is now genuinely competitive with a 200B+ closed model at $3 per million tokens.
The frontier is no longer where the largest model lives. It is where the most useful model lives for what you are trying to ship. Meta just re-staked their claim on the right side of that line. You ship it. They keep the open-source throne warm. Closed-source labs keep selling to the workloads that need 200B+ parameters. Everyone does what they are good at.
— Mr. Technology
Sources: