WebBrain Compass Tiny v2: small model, sharper tool routing
Meet WebBrain Compass Tiny v2, our new small model for choosing WebBrain browser tools and their arguments. Built on MiniCPM5-2B, it shows a substantial routing improvement over its base. We expect to incorporate it into WebBrain soon, subject to runtime validation and release checks.
A focused upgrade
Compass Tiny v2 is a LoRA fine-tune of OpenBMB's MiniCPM5-2B, approximately 2.6B parameters. We trained it for one epoch on 7,311 reviewed examples, with 213 held out for validation. Its job is specialized tool routing—not replacing a general-purpose planner.
Against the separately tested base, the merged BF16 model moves from 9 to 16 exact actions and 21 to 41 tool-family matches. Structured-call coverage slips from 96 to 94 out of 100.
One comparison, all models
Sorted by strict exact action, then loose match. Larger Qwen models are included as reference points. Both bar columns use the same 0–89 scale.
| Model | Parameters | Structured calls | Strict exact action | Loose tool-family match | Loose rate |
|---|---|---|---|---|---|
| Qwen3.8-27B | 27B | 99 / 100 | 17 / 89 | 39 / 89 | 43.8% |
| WebBrain Compass Tiny v2Merged BF16 · new | ~2.6B | 94 / 100 | 16 / 89 | 41 / 89 | 46.1% |
| Qwen3.5-4B | 4B | 89 / 100 | 16 / 89 | 36 / 89 | 40.4% |
| Gemma 4 E4B | 7.5B | 90 / 100 | 15 / 89 | 43 / 89 | 48.3% |
| Qwen3.6-35B-A3B | 35B A3B | 89 / 100 | 15 / 89 | 42 / 89 | 47.2% |
| Nanbeige4.2-3B | 4.2B reported | 89 / 100 | 13 / 89 | 37 / 89 | 41.6% |
| Gemma 4 12B QAT | 12B | 90 / 100 | 12 / 89 | 34 / 89 | 38.2% |
| Qwen3.5-2B | 2B | 92 / 100 | 11 / 89 | 28 / 89 | 31.5% |
| MiniCPM5-2B baseCompass Tiny v2's base | ~2.6B | 96 / 100 | 9 / 89 | 21 / 89 | 23.6% |
| Gemma 4 E2B | 4.6B | 74 / 100 | 8 / 89 | 45 / 89 | 50.6% |
| WebBrain Compass Tiny v1 | 2.6B | 81 / 100 | 4 / 89 | 42 / 89 | 47.2% |
| LFM2.5-2.6B base | 2.6B | 80 / 100 | 4 / 89 | 34 / 89 | 38.2% |
Compass Tiny v2 is one exact match behind Qwen3.8-27B and two tool-family matches ahead in these runs—a promising result at roughly one-tenth the parameter count, not a claim of general 27B-level capability.
What's next
The BF16 research preview is private on Hugging Face. ONNX compatibility work is underway; integration is planned, not shipped. Runtime checks and training-data licensing clearance remain prerequisites.
Results combine our earlier comparison, the author's separate MiniCPM5 base run, and Compass Tiny v2's merged BF16 run and saved results. Backends and precision differ; this is not a controlled speed comparison. Browser actions were not executed. Compass Tiny v2 retained one malformed-XML response and missed our earlier adapter-relative loose-match target (41 versus 43), so it remains experimental.
Tags: #WebBrain #CompassTinyV2 #MiniCPM5 #ToolCalling #LocalAI