14B-32B local models — all 47 · V1
Local Model Testing · Tier v3.0
The third Sigma5C local-model observation. Forty-one local models in the 9B–32B range plus six cloud-routed models. The same v2.0 six-category framework. CPU-only inference on a 48-thread Xeon CPU with 157 GB RAM (Ollama). This tier is where the practical CPU ceiling lives — 14B is the sweet spot at ~2.7-3.0 tok/s; 32B is unusable on CPU.
- top (qwen3-coder:30b)
- 9.53
- local avg
- 6.54
- cloud avg
- 9.01
- total (47 + 6 Qwen 3.5)
- 53
What this observation taught us
Champion
qwen3-coder:30b — 9.53 / 10
Wins on every category except Reason (9.12) where it sits second. Highest-scoring CPU-only local model observed across all three Sigma5C local-model tiers. 10.3 tok/s on the Xeon CPU.
CPU sweet spot
14B is usable, 32B is not
14B runs at ~2.7–3.0 tok/s — workable. 32B models slow to 1.5–2.0 tok/s on CPU and frequently time out on hard prompts.
The full 14B-32B observed score table
- qwen3-coder:30b
- 9.53
- ministral-3:14b-cloud
- 9.31
- gemma3:27b-cloud
- 9.3
- devstral-small-2:24b-cloud
- 9.26
- gemma2:27b
- 9.18
- gemma3n:latest
- 9.04
- nemotron-3-nano:30b-cloud
- 8.99
- cogito:14b
- 8.86
- gemma3:12b-cloud
- 8.81
- qwen2.5-coder:32b
- 8.6
- qwen2.5vl:latest
- 8.59
- qwen2.5:32b
- 8.56
- qwen2.5-coder-14b-optimized:latest
- 8.54
- qwen2.5:14b
- 8.48
- gemma2:latest
- 8.44
- gemma2:9b
- 8.43
- gpt-oss:20b-cloud
- 8.38
- qwen2.5-coder-32b-optimized:latest
- 8.35
- gpt-oss-safeguard:20b
- 8.18
- falcon3:10b
- 8.01
- devstral:24b
- 7.94
- qwen2.5-coder:14b
- 7.94
- granite4:small-h
- 7.93
- gpt-oss:20b
- 7.84
- cogito:32b
- 7.73
- mistral-small3.2:24b
- 7.65
- ministral-3:14b
- 7.62
- qwen:32b
- 7.6
- devstral-small-2:latest
- 7.52
- nemotron-3-nano:30b
- 7.49
- devstral:latest
- 7.45
- mistral-small3.2:latest
- 7.11
- dolphin-mistral-nemo:12b
- 6.34
- qwen3:30b
- 6.19
- mistral-nemo-uncensored:latest
- 6.11
- gemma3:12b
- 5.78
- qwen2.5vl:32b
- 5.58
- gemma3:27b
- 5.38
- qwen:14b
- 5.33
- deepseek-ocr:latest
- 2.7
- deepcoder:14b
- 2.35
- qwen3-vl:latest
- 2.28
- magistral:24b
- 1.98
- qwen3:14b
- 1.95
- qwen3:32b
- 1.85
- srizon/pixie:latest
- 1.2
- exaone-deep:32b
- 0.5
47 entries · ordered by Overall as observed in this run · not a ranking
Six Qwen 3.5 models added — 53 total
When the Qwen 3.5 family shipped, six were run through the same v3.0 suite, taking the tier to 53 models observed. On these deliberately hard 14B-32B prompts the smaller Qwen 3.5 sizes struggled: qwen3.5:2b led the addendum at 4.38 with strong Code (8.10) but near-zero Reasoning and Math, and qwen3.5:27b did not complete — it errored on every prompt.
- qwen3.5:2b
- 4.38
- qwen3.5:0.8b
- 2.99
- qwopus-9b:latest
- 2.88
- qwen3.5:4b
- 2.78
- qwen3.5:latest
- 1.65
- qwen3.5:27b
- 0
6 entries · ordered by Overall as observed in this run · not a ranking
Models
- cogito:14b
- cogito:32b
- deepcoder:14b
- deepseek-ocr:latest
- devstral:24b
- devstral:latest
- devstral-small-2:24b-cloud
- devstral-small-2:latest
- dolphin-mistral-nemo:12b
- exaone-deep:32b
- falcon3:10b
- gemma2:27b
- gemma2:9b
- gemma2:latest
- gemma3:12b-cloud
- gemma3:12b
- gemma3:27b-cloud
- gemma3:27b
- gemma3n:latest
- gpt-oss:20b-cloud
- gpt-oss:20b
- gpt-oss-safeguard:20b
- granite4:small-h
- magistral:24b
- ministral-3:14b-cloud
- ministral-3:14b
- mistral-nemo-uncensored:latest
- mistral-small3.2:24b
- mistral-small3.2:latest
- nemotron-3-nano:30b-cloud
- nemotron-3-nano:30b
- qwen:14b
- qwen:32b
- qwen2.5:14b
- qwen2.5:32b
- qwen2.5-coder-14b-optimized:latest
- qwen2.5-coder:14b
- qwen2.5-coder-32b-optimized:latest
- qwen2.5-coder:32b
- qwen2.5vl:32b
- qwen2.5vl:latest
- qwen3:14b
- qwen3:30b
- qwen3:32b
- qwen3.5:0.8b
- qwen3-coder:30b
- qwen3-vl:latest
- srizon/pixie:latest