Meta Muse Glimmer 30B: First Open-Weight Model Built for Always-On Local Agent Workflows on a Single RTX 5090

Overview

On August 10, Meta Superintelligence Labs and Scale AI jointly released Muse Glimmer, a 30B-parameter multimodal dense model under the Apache 2.0 license. Unlike prior open-weight large models optimized for leaderboard scores, Muse Glimmer is explicitly engineered for always-on, local agent workflows. The same day, the LMSYS SGLang team added Day-0 support, reducing deployment to a single RTX 5090.

Architecture Highlights

• 30B total = 27.9B text decoder + 1.9B ViT + GELU multimodal projector. The visual encoder is kept separate so it does not consume text-side capacity.
• Text decoder: 52 transformer layers, grouped-query attention (32 query heads / 2 KV heads), SwiGLU FFN.
• Hybrid attention pattern: alternating 3 × 2048-token sliding-window layers with 1 full-sequence attention layer. Sliding-window layers use RoPE; full-attention layers use NoPE, allowing context length to extend beyond the training window.
• 128k+ token context window — sufficient for 4–5 full repositories plus MCP tool protocols plus multi-turn dialogue history, without summarization or compression.
• Agent-tuned training: LMSYS reports the model remains competitive with leading models of similar size on key agent benchmarks (full numbers pending the Meta paper).

SGLang Day-0 Optimization Stack

• Throughput on RTX 5090 (NVFP4 + DFlash speculative decoding): 236 tok/s/user at batch 1, 1452 tok/s aggregate at batch 8. DFlash delivers a 1.9–4.3× speedup (236 / 63.9 ≈ 3.7×).
• Multi-hardware paths:
• NVFP4 — NVIDIA SM120 backend (RTX 5090, RTX PRO 6000, DGX Spark)
• BF16 — single H100
• GGUF Q4KM (18 GB) / Q4K-Dynamic — 24 GB consumer cards
• MLX — Apple Silicon Mac mini / M-series MacBook Pro (56.9 tok/s Q4K-Dynamic, batch 1 on M5 Pro)
• Three native optimizations: low-overhead scheduler, RadixAttention prefix cache, interruptible CUDA graphs. Prefix cache is critical for agents: repeated tool schemas, recurring code reviews, and multi-turn dialogue hit identical prefixes at zero cost.
• DFlash speculative decoding: draft and target models are decoupled; the target model only verifies multi-token candidates, doubling throughput without quality loss.

Why This Matters for AI Coding Tooling

• 12-month baseline: Claude Sonnet, GPT-5, and Gemini 2.5 Pro dominated cloud inference. Local options were limited to "barely-usable" 7B-class models (Qwen2.5-Coder-7B, Code Llama 7B, DeepSeek Coder 6.7B). True local agents were impractical.
• 6-month progress: 32B-class open agent models emerged (Qwen3.5-32B, DeepSeek-V3-Flash, GLM-4.5-Air) — but none were designed for 24/7 residency, treating peak memory, long-context KV cache, frequent tool-schema calls, and prefix-cache hit rate as afterthoughts.
• Muse Glimmer's answer: 30B + 128k+ context + 24 GB footprint + Apache 2.0 + SGLang Day-0 + DFlash + RadixAttention + MLX cross-hardware support — the first open-weight model born for local agents.

Concrete Implications

• Viable local agent backend: Previously, 70B models required 2–4 H100s or 8–16 RTX 4090s; 7B models lacked capability. Muse Glimmer at 236 tok/s on a single RTX 5090 gives harnesses like Claude Code, Codex, and OpenCode their first locally runnable and sufficiently capable backend option.
• Privacy/compliance unlocked: Legal, medical, financial, government, and consulting workflows with "code must not leave the company machine" requirements previously had to settle for 7B compromise models or self-hosted data centers. Muse Glimmer compresses this path down to a single RTX 5090 workstation.
• Cost structure reshaped: Claude Sonnet API pricing of $3/$15 per 1M tokens means a 200k-token overnight agent run costs ~$3. Running Muse Glimmer locally incurs only electricity (RTX 5090 ≈ 400 W), enabling "re-run the full test suite every night" workloads without API-bill anxiety.

Notable Details

• Scale AI co-credit: Alex Wang confirmed on August 10 that Muse Glimmer 30B and Muse Spark 1.2 open-sourced together. Scale contributed data/RLHF pipelines; Meta contributed architecture/training. Scale AI's RLHF training moat is now publicly visible.
• Apache 2.0 is pivotal: Llama's Community License restricted products exceeding 700M MAU. Apache 2.0 removes that friction for AI coding vendors (Cursor, Devin, Replit Agent, OpenChamber, etc.) — the license risk for a local backend is eliminated for the first time.
• SGLang's moat: vLLM can also serve 30B models, but SGLang's integrated prefix-cache + speculative-decoding + multi-hardware stack is what captures the 30B-era upside.
• Apple Silicon path: 56.9 tok/s on M5 Pro means MacBook Pro users can run a 30B agent locally — a segment previously limited to 7B compromise models.

August 2026 AI Coding Timeline Context

• 08-04 — Ant Ling-3.0-flash open-sources (KDA+MLA + 1/64 sparse MoE + SGLang HiCache + Mooncake).
• 08-05 — NVIDIA Cosmos 3 (64B / 16B / 4B embodied world models).
• 08-09 — Microsoft SkillOpt (portable experience layer).
• 08-10 — Meta Muse Glimmer 30B — the first open-weight model designed for always-on local agents.

August 10 marks the formal opening of the localization era for AI coding: developers finally have a genuinely comparable choice between cloud Sonnet/Opus and local 30B, rather than cloud versus a compromised 7B.

Sources

• LMSYS SGLang blog: https://www.lmsys.org/blog/2026-08-10-meta-muse-glimmer
• Meta @AIatMeta announcement: https://x.com/AIatMeta/status/2086757844544811485
• Scale AI @alexandr_wang announcement: https://x.com/alexandr_wang/status/2086756152034066792
• Meta Research details: https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model
• SGLang installation: https://github.com/sgl-project/sglang
• DFlash speculative decoding: https://github.com/sgl-project/sglang/tree/main/python/sglang/speculative