MCP Atlas (LFM3 report, unspecified protocol)
Agents
MCP-Atlas pass rates that Liquid AI ran for its LFM3 launch comparison, with up to 256 turns per task. The card does not establish the task subset, judge or its revision, pass threshold or retry handling, and its two-decimal values match a single run over neither the 500-task public set nor the 1,000-task leaderboard set. Retained as a source-specific diagnostic.
Top models (higher is better)
| Model | Score |
|---|---|
| Qwen3.5-9B | 42.7 |
| Qwen3.5-4B | 41.4 |
| LFM3-1.7B | 32.6 |
| LFM2.5-2.6B | 16.4 |