3B-7B local models — all 53 · V1
Local Model Testing · Tier v2.0
- avg overall
- 6.91
- local avg
- 6.70
- cloud avg
- 8.92
- cloud premium
- +2.22
- v1→v2 drop
- -2.06
Harder prompts, an Analysis category, broader model lineup
Of the nine models scored under both v1.0 and v2.0 prompts, the average drop was -2.06 points (v1.0 avg 9.36 → v2.0 avg 7.29). The largest single drop was phi3.5:latest at -2.51. This validates the v2.0 prompt design: harder prompts have separated the leaders from the followers.
Headline findings
#1 overall
qwen2.5:7b
9.71 / 10. Highest-scoring local model. Wins reasoning, instruction, summary, math, and analysis categories outright.
Cloud calibration
ministral-3:8b-cloud
9.61. Highest-scoring cloud-routed entry. Tier-A wraparound — close to qwen2.5:7b despite being remote.
Surprise small
granite4:3b
8.48 at 12.6 tok/s. Strong perf-per-byte. Fastest local: granite3-moe:3b at 31.7 tok/s.
Hardest category: Analysis (avg 5.21 / 10). The new race-condition and complexity reasoning prompts separated the top tier (9+ on Analysis) from the rest. Easiest category: Instruction Following (avg 8.30 / 10).
The full 3B-7B observed score table
-cloud in the model name; their tok/s appears as "cloud" since CPU-throughput is not applicable.- qwen2.5:7b
- 9.71
- ministral-3:8b-cloud
- 9.61
- gpt-oss:20b-cloud
- 9.4
- exaone3.5:latest
- 8.89
- qwen2.5-coder:7b
- 8.86
- gemma3n:e4b
- 8.81
- translategemma:4b
- 8.76
- falcon3:7b
- 8.69
- ministral-3:3b-cloud
- 8.68
- gemma3:4b-cloud
- 8.63
- ministral-3:3b
- 8.62
- command-r7b:latest
- 8.59
- granite4:3b
- 8.48
- granite3.3:8b
- 8.46
- rnj-1:8b-cloud
- 8.26
- gemma3:4b
- 8.26
- qwen2.5vl:7b
- 8.25
- qwen2:7b
- 8.21
- qwen2.5-coder-sigma5c:7b
- 8.19
- cogito:3b
- 7.99
- dolphin3:latest
- 7.92
- huihui_ai/dolphin3-abliterated:8b
- 7.88
- cogito:8b
- 7.81
- granite4:latest
- 7.66
- dolphin-2.9.2-qwen2-7b:Q3_K_M
- 7.63
- qwen2-math:latest
- 7.54
- phi4-mini:latest
- 7.44
- qwen2.5:3b
- 7.43
- phi3:latest
- 7.4
- llama3.2:3b
- 7.4
- falcon3:3b
- 7.36
- qwen2.5-coder:3b
- 7.35
- lfm2.5-thinking:latest
- 7.28
- dolphin-mistral:latest
- 7.19
- dolphin-llama3.1:latest
- 7.09
- granite3.1-moe:3b
- 6.93
- olmo2:7b
- 6.89
- phi3.5:latest
- 6.66
- hermes3:3b
- 6.6
- codeqwen:latest
- 5.87
- qwen:7b
- 5.49
- granite3-moe:3b
- 5.46
- qwen3-vl:4b
- 4.89
- dolphin-2.9.2-qwen2-7b:Q2_K
- 4.53
- deepseek-r1:7b
- 4.15
- deepseek-r1:8b
- 4.01
- qwen3:4b
- 3.8
- deepscaler:latest
- 3.75
- qwen:4b
- 3.26
- openthinker:7b
- 2.2
- smallthinker:3b
- 2.05
- qwen3:8b
- 1.9
- exaone-deep:7.8b
- 1.88
53 entries · ordered by Overall as observed in this run · not a ranking
Models
- codeqwen:latest
- cogito:3b
- cogito:8b
- command-r7b:latest
- deepscaler:latest
- deepseek-r1:7b
- deepseek-r1:8b
- dolphin-2.9.2-qwen2-7b:Q2_K
- dolphin-2.9.2-qwen2-7b:Q3_K_M
- dolphin-llama3.1:latest
- dolphin-mistral:latest
- dolphin3:latest
- exaone-deep:7.8b
- exaone3.5:latest
- falcon3:3b
- falcon3:7b
- gemma3:4b-cloud
- gemma3:4b
- gemma3n:e4b
- gpt-oss:20b-cloud
- granite3.1-moe:3b
- granite3.3:8b
- granite3-moe:3b
- granite4:3b
- granite4:latest
- hermes3:3b
- huihui_ai/dolphin3-abliterated:8b
- lfm2.5-thinking:latest
- llama3.2:3b
- ministral-3:3b-cloud
- ministral-3:3b
- ministral-3:8b-cloud
- olmo2:7b
- openthinker:7b
- phi3.5:latest
- phi3:latest
- phi4-mini:latest
- qwen:4b
- qwen:7b
- qwen2.5:3b
- qwen2.5:7b
- qwen2.5-coder:3b
- qwen2.5-coder:7b
- qwen2.5-coder-sigma5c:7b
- qwen2.5vl:7b
- qwen2:7b
- qwen2-math:latest
- qwen3:4b
- qwen3:8b
- qwen3-vl:4b
- rnj-1:8b-cloud
- smallthinker:3b
- translategemma:4b