Meta dropped Muse Glimmer on August 10 — a 30B open-weight model built for one purpose: running a capable coding agent on hardware you already own. Download it from Hugging Face, load it in Ollama, and point it at your codebase. Nothing leaves your machine. No API bill at the end of the month.
That sounds like a modest promise until you consider what “hardware you already own” means in August 2026. With its 24GB quantized build, Muse Glimmer fits on an M4 Pro MacBook, an RTX 4090, or an AMD 7900XT. Developers in the Hacker News thread confirmed it loads at 16.76GB for the language model plus 1.40GB for the vision encoder. It works. It is not theoretical.
What Muse Glimmer Actually Is #
Muse Glimmer is a distillation of Muse Spark 1.2 — Meta’s closed frontier model — compressed into a 30B dense transformer with a 131K-token context window, text and image inputs, and an Apache 2.0 license that allows commercial fine-tuning. Meta trained it specifically for agentic workloads: multi-step tool calling, failure recovery, and autonomous code editing across sessions that run for hours.
Two quantization builds ship at launch. The K-Quant-17GB targets 24GB VRAM at a reported 1.0% benchmark degradation. The K-Quant-Dynamic targets 32GB at 0.2% degradation. Full precision requires 55GB — the quantized builds are the real product here. Pair them with the DFlash speculative decoder and you get a 3.1x throughput boost on an RTX 5090, or a 1.6–1.9x boost on Apple Silicon. Community members on an AMD 7900XT measured roughly 36 tokens per second without DFlash, jumping to 60 tokens per second with it enabled.
It integrates with Ollama, LM Studio, llama.cpp, MLX, and vLLM from day one. Getting it running is one command:
ollama pull muse-glimmer:30b-q4
Muse Glimmer vs Qwen 3.6: Pick Your Problem #
The obvious comparison is Qwen 3.6 27B, the dominant local model heading into August. The results are messier than Meta’s press release implies.
On agentic task reliability — the stuff that matters for a coding agent — Muse Glimmer is clearly ahead. It passed or partially passed five of seven tasks in a community benchmark (3 full passes, 2 partials); Qwen managed two passes and five failures. Qwen’s critical failure mode was consistent: write_file
calls returned unterminated JSON, triggering HTTP 500 errors. Muse’s failure mode was quieter — it would plan multi-file edits and then stop without writing anything. Still a failure, but not a hard crash.
On speed, Qwen wins decisively. Its mixture-of-experts architecture activates only about 3B parameters per token. Muse Glimmer activates all 32.3B. That dense-versus-sparse tradeoff makes Qwen roughly 2x faster — 15.6 seconds average per task versus 31 seconds. For agents running hundreds of sequential tasks, that gap compounds.
There is also the hardware reality: Qwen 3.6 27B ran out of memory on the same 24GB Mac during the benchmark. It needed an A100 to complete the test. If you have an M4 Pro MacBook or a 24GB consumer GPU, the comparison is effectively moot — Muse Glimmer is the only option that fits.
| Model | MCP Atlas | TerminalBench 2.1 | SWE-Bench Verified | Fits 24GB? |
|---|---|---|---|---|
| Muse Glimmer 30B | 75.5 | 54.2 | 65.9 | Yes |
| Qwen 3.6 27B | 62.5 | 60.7 | 77.2 | No (OOM) |
| Gemma 4 31B | 54.2 | — | — | Yes |
On raw coding benchmarks — TerminalBench and SWE-Bench Verified — Qwen leads by meaningful margins when it has sufficient VRAM. Muse Glimmer dominates on MCP Atlas, which measures multi-tool orchestration: exactly what a coding agent spends most of its time doing.
What It Cannot Do Yet #
Multi-file editing is still unreliable. Both models failed to complete edits spanning multiple files during community testing — the most common real-world coding agent task. Neither is production-ready for complex refactors without a human approval loop at each step.
The HN community also pushed back on the “open source” framing. Meta released weights under Apache 2.0 — commercially useful and genuinely permissive — but there is no training data, no training code, and no reproducibility pipeline. Open weights on Hugging Face are not the same thing as open source. Developers evaluating this for enterprise compliance should understand the distinction before signing off on a deployment.
The Verdict #
Muse Glimmer is the best local coding agent model available right now for developers on 24GB consumer hardware — and that hardware bracket covers most serious developers working outside a server room. The tool-calling reliability is real; one HN developer described its reasoning traces as “a cocky bastard in tone, but relatively terse” — which is about the best thing you can say about an agentic model’s personality.
What it is not: a replacement for cloud agents on complex multi-file work. Use Muse Glimmer for the tasks where local execution matters most — privacy-sensitive analysis, overnight batch jobs, air-gapped environments — and keep a cloud agent in the workflow for heavy refactors. That is not a failure state. It is a practical division of labor that did not exist six months ago.