Run 1
4:30 min:stime
11,710output tokens
2.4%thinking
43.43tok/s
624lines of code
Effort requested: medium; received: on · differs from requested effort · source: metrics.json. Effort is not comparable across models.
qwen3.6-35b-a3b_bare_r1
Qwen 3.6 35B A3B puts 70+ effects on every page. Do rules tone it down?
qwen/qwen3.6-35b-a3b
Same model, same prompt, same rig. 3 runs with nothing added, 3 with one Karpathy rules file.
qwen/qwen3.6-35b-a3b
Ranges overlap (counted effects)Effects count CSS/JS constructs, not visual quality.
Repeated ranges
70–92
KARPATHY · n=377–87
4:30–4:47 min:s
KARPATHY4:23–4:39 min:s
11,710–12,607
KARPATHY11,588–12,198
43.43–44.21 tok/s
KARPATHY43.56–44.09 tok/s
2.4%–3.8%
KARPATHY2.2%–4.1%
579–624
KARPATHY561–615
BARENO RULES, NO EXTRAS
4:39 min:stime
12,327output tokens
3.8%thinking
579lines of code
Effort requested: medium; received: on · differs from requested effort · source: metrics.json. Effort is not comparable across models.
KARPATHYONE RULES FILE
4:37 min:stime
12,056output tokens
2.2%thinking
615lines of code
Effort requested: medium; received: on · differs from requested effort · source: metrics.json. Effort is not comparable across models.
Every run. Every result.
BARE
NO RULES, NO EXTRAS · n=3
Run 2
4:47 min:stime
12,607output tokens
3.8%thinking
43.92tok/s
601lines of code
Effort requested: medium; received: on · differs from requested effort · source: metrics.json. Effort is not comparable across models.
qwen3.6-35b-a3b_bare_r2
Run 3
4:39 min:stime
12,327output tokens
3.8%thinking
44.21tok/s
579lines of code
Effort requested: medium; received: on · differs from requested effort · source: metrics.json. Effort is not comparable across models.
qwen3.6-35b-a3b_bare_r3
KARPATHY
ONE RULES FILE · n=3Run 1
84effects
4:37 min:stime
12,056output tokens
2.2%thinking
43.56tok/s
615lines of code
Effort requested: medium; received: on · differs from requested effort · source: metrics.json. Effort is not comparable across models.
qwen3.6-35b-a3b_karpathy_r1
Run 2
77effects
4:39 min:stime
12,198output tokens
2.8%thinking
43.67tok/s
561lines of code
Effort requested: medium; received: on · differs from requested effort · source: metrics.json. Effort is not comparable across models.
qwen3.6-35b-a3b_karpathy_r2
Run 3
87effects
4:23 min:stime
11,588output tokens
4.1%thinking
44.09tok/s
579lines of code
Effort requested: medium; received: on · differs from requested effort · source: metrics.json. Effort is not comparable across models.
qwen3.6-35b-a3b_karpathy_r3
output = thinking + final code. Lines of code show the page size the model chose, not the cost of equivalent work.
How to reproduce
Settings recorded for these runs. Multiple values show differences between runs. Effort is requested, not necessarily received; it is not comparable across models. Hover over quantization, K/V cache or engine values for their sources.
qwen/qwen3.6-35b-a3b
| Model | Quantization | Context window | K/V cache | max_tokens | Engine | Harness | Effort requested |
|---|---|---|---|---|---|---|---|
| qwen/qwen3.6-35b-a3b | Q4_K_M | 262,144 | K q8_0 / V q8_0 | 131,072 | [email protected] | pi 0.85.1 | medium |
Watch the experiment.
REPRODUCE THE EXPERIMENT
One prompt.
All the evidence.
Compare runs within the same model. Inspect the shared prompt, settings and raw measurements in the repository.