LOCAL INFERENCE / EIGHT TASKS
GLM vs Qwen.
The same eight tasks.
We gave identical prompts to two models on the same pair of RTX 3090s, one after the other. Every answer was run in a real browser, then judged blind by Claude. The question: which local model actually ships working work, and how fast?
4GLM-5.3-Flash
4Qwen3.8-27B
A tie. GLM is more polished when it works, but 4 of its 8 outputs crashed. All 8 of Qwen's worked, in 5m 19s total.
The numbers
Average speed
GLM16 tok/s
Qwen187 tok/s
The models
GLM-5.3-Flash
- Format
- EXL3 2.05 bpw, experts tiered across VRAM, RAM and NVMe
- Profile
- faithful
- Sampling
- temperature 1.0, top-p 0.95
- Thinking
- on, 10,240-token budget
- Output cap
- 32,768 tokens
Qwen3.8-27B
- Format
- EXL3 with DFlash2, Elpis engine, two lanes
- Sampling
- greedy (the only mode the engine supports)
- Thinking
- on, effort medium
- Output cap
- 32,768 tokens
Hardware: 2 × RTX 3090, 31 GB RAM. The two models never ran at the same time.
The tasks
Pelican on a bicycle
SVG animation
GLM-5.3-Flash27m 57s · 32,768 tokens · 20 tok/s
Qwen3.8-27B1m 01s · 11,046 tokens · 181 tok/s
GLM-5.3-Flash · Richer scene, more detailed pelican and bike.
Twenty balls in a spinning hexagon
Physics
GLM-5.3-Flash20m 46s · 16,620 tokens · 13 tok/s
Qwen3.8-27B29s · 5,541 tokens · 188 tok/s
GLM-5.3-Flash · Polished interface, energy graph, grab and throw.
Solar system
Three.js
GLM-5.3-Flash23m 07s · 21,618 tokens · 16 tok/s
Qwen3.8-27B28s · 5,311 tokens · 191 tok/s
Qwen3.8-27B · Orbits and ringed planet work. GLM's page stayed stuck on its loading screen.
Island with a day cycle
Three.js
GLM-5.3-Flash29m 36s · 32,768 tokens · 18 tok/s
Qwen3.8-27B40s · 7,209 tokens · 178 tok/s
Qwen3.8-27B · Island renders with a campfire and day cycle. GLM's code had a syntax error and drew nothing.
Snake game
Game
GLM-5.3-Flash20m 35s · 18,441 tokens · 15 tok/s
Qwen3.8-27B57s · 10,623 tokens · 185 tok/s
Qwen3.8-27B · Playable. GLM's board stayed empty after a negative-radius error.
Coffee subscription landing page
Landing page
GLM-5.3-Flash23m 07s · 22,281 tokens · 16 tok/s
Qwen3.8-27B48s · 9,462 tokens · 198 tok/s
GLM-5.3-Flash · Both are good. GLM built more sections and a fuller page.
Double pendulum
Simulation
GLM-5.3-Flash22m 02s · 18,306 tokens · 14 tok/s
Qwen3.8-27B33s · 6,203 tokens · 187 tok/s
GLM-5.3-Flash · Complete controls, presets and a divergence graph.
Spiral galaxy
WebGL shader
GLM-5.3-Flash20m 34s · 15,071 tokens · 12 tok/s
Qwen3.8-27B22s · 4,233 tokens · 192 tok/s
Qwen3.8-27B · Clear spiral arms. GLM's shader failed to compile, leaving a blurry blob.
Honest notes
- Each task ran once. This is a first impression, not a ranking.
- Claude judged. Screenshots were shown in random A/B order with model names hidden, taken after scripted keyboard and mouse input.
- Qwen was first tried at effort high. In 7 of 8 tasks it spent all 32,768 tokens thinking, and 4 never produced an answer. It was rerun at medium, which is what you see here.
- GLM got a 10,240-token thinking budget and Qwen has no such limit. GLM's budget forcing adds time: the server shows about 35 tok/s, but end to end it averaged 16 tok/s.
- Both ran on the same hardware but work very differently. One sits fully in VRAM, the other pulls experts from RAM and disk.