Muse Glimmer 30B takes way longer to think than Qwopus 27B for

MaxCrafter Novice 1h ago 198 views 8 likes 2 min read

I just wrapped up a head-to-head comparison between Qwopus 3.6 27B and Muse Glimmer 30B using my own production code instead of some sanitized benchmark dataset. I ran both locally on an AMD Radeon RX 7900 XT (20GB VRAM) via llama.cpp. To keep things objective, I used a third model as a referee to review the diffs and relied on a strict Vitest suite to verify if the code actually worked.

For the setup, I used GGUF quants (IQ4_XS for Qwopus and UD-Q4_K_XL for Muse). Both were pushed through a pi CLI with --thinking high. The project was a vanilla JS PWA called "Jeu de Cochons"—real-world code with actual deployment constraints.

The Regression Test

The first task was a surgical strike: fix three broken service-worker tests caused by a base path change for GitHub Pages. The constraint was simple—fix the bugs without touching the test files.

Qwopus 3.6 27B: 11/11 PWA tests passed, 87/87 full suite passed. Time: ~8.5 min.
Muse Glimmer 30B: 11/11 PWA tests passed, 87/87 full suite passed. Time: ~21 min.

The result was wild—both models produced a byte-identical diff. They both identified the exact same regex capture group error. However, Muse took 2.5x longer to arrive at the exact same conclusion. Interestingly, the referee model pointed out that while both fixed the tests, they both missed a latent production conflict where reverting the manifest to root would cause 404s on the live site. This tells me that for narrow, well-defined bugs, the "thinking" overhead of the larger model might be wasted effort.

Feature Implementation Deep Dive

To really stress-test their reasoning, I asked them to implement a single-player mode against an AI opponent. This required updating the game state, creating a bot decision strategy, and writing new integration tests.

Qwopus 27B Performance:

- New tests: 18 written (12 unit, 6 integration).
- Success rate: 102/105 tests passed.
- Logic: Developed an adaptive strategy (banking at 25 pts normally, but dropping to 15 pts if leading).
- Cleanliness: No leftover artifacts.

Muse Glimmer 30B Performance:

- New tests: 7 written.
- Success rate: 91/94 tests passed.
- Logic: Used a fixed threshold of 15 pts.
- Cleanliness: Left uncommitted dist/ folders in the workspace.

Qwopus didn't just write more tests; it actually designed a more "human" AI strategy that adjusted based on the game score. Muse was more rigid and slightly sloppier with the file system.

Final Verdict

If you're looking for a local LLM agent for a real-world AI workflow, Qwopus 27B feels like the more efficient choice. It matches or beats Muse Glimmer 30B in logic and test coverage while being significantly faster. Muse Glimmer is capable, but the latency-to-quality ratio isn't favoring it here. For those building a local deployment, the 27B parameter class seems to be the current sweet spot for coding tasks.