September 7, 2026 · 3 min read · ← All posts

Compact tool routing compared: 10 models, from 2B to 35B

We added six local models to the WebBrain Compact routing benchmark. The updated table now spans 10 models from 2B to 35B. Qwen3.8-27B leads strict exact-action accuracy and first-turn structured-call coverage. A separate, looser tool-family measure produces a different ordering, but should not be read as end-to-end scenario success.

Updated results

The suite contains 100 first-turn prompts and 100 stateful scenarios. Eleven scenarios are intentionally outside Compact mode, leaving 89 scored scenarios. Strict exact action requires both the reference tool name and its arguments to match. Loose tool-family match also gives credit for name-only matches and terminal prose mapped to the expected tool family; it is diagnostic, not a task-success score.

ModelParametersFirst-turn structured callsStrict exact action (/89)Loose tool-family match (/89)Loose match rate
Qwen3.8-27B27B99 / 100173943.8%
Qwen3.5-4B4B89 / 100163640.4%
Qwen3.6-35B-A3B35B A3B89 / 100154247.2%
Gemma 4 E4B7.5B90 / 100154348.3%
Nanbeige4.2-3B4.2B reported89 / 100133741.6%
Gemma 4 12B QAT12B90 / 100123438.2%
Qwen3.5-2B2B92 / 100112831.5%
Gemma 4 E2B4.6B74 / 10084550.6%
WebBrain Compass Tiny v12.6B81 / 10044247.2%
LFM2.5-2.6B base2.6B80 / 10043438.2%

What stands out

Notes

The six new rows—Gemma 4 E2B, Gemma 4 E4B, Gemma 4 12B QAT, Nanbeige4.2-3B, Qwen3.5-4B, and Qwen3.5-2B—were run locally as 4-bit models through LM Studio. Concurrency varied from one to four, so this is a quality table, not a speed comparison. The displayed E4B row is its second concurrency-2 run; the first produced 13 exact and 42 loose matches, illustrating modest run-to-run variance.

The runner records the model’s next response without executing the browser action. These are routing results, not end-to-end task-completion rates. The Sonnet-authored reference action is a regression target rather than an oracle: a cautious model may reasonably inspect before acting. In the paired E2B/Qwen3.8 comparison, all 19 scenarios credited only to E2B were loose ideal_name outcomes, not exact action-and-argument matches.

Tags: #WebBrain #BrowserAgent #ToolCalling #CompactMode #Gemma4 #Qwen #CompassTiny

Written by Emre Sokullu. WebBrain 33.0.0 and later is GPL-3.0-or-later and open on GitHub.