LFM3-1.7B Release - τ³-bench (4 domains), average: leaderboard
Liquid AI LFM 3 text release card; vendor reruns in vLLM BF 16 at each full context length (262144 for LFM 3 and Qwen 3.5; 131072 for other models). Default temperature 0.6 and top_p 0.95. BFCL V 4 greedy/max 32768; ToolSandbox greedy/max 1024; LiveClaw OpenClaw/max 8192; Claw-Eval greedy/3 trials; BrowseComp-Plus OpenClaw/greedy/max 8192; WebArena max 40 steps; MCP-Atlas max 256 turns; EnterpriseOps/Automation max 50 turns; SWE-bench OpenCode/max 250 turns; Terminal-Bench 2.1 OpenCode/max 100 turns. Context is the output limit otherwise. No additional reasoning effort is inferred. Metric: Score (%).
Metric: Score (%). Source: huggingface.co. Saturation forecast: Around February 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | MiniCPM5-2B | 69.44 | #596 |
| 2 | LFM3-1.7B | 68.01 | #501 |
| 3 | Qwen 3.5 9B | 67.11 | #391 |
| 4 | Qwen 3.5 4B | 65.98 | #514 |
| 5 | LFM2.5-2.6B | 56.77 | #836 |
| 6 | Gemma 4 E4B | 33.53 | #762 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=lfm3-1-7b-release-tau3-bench-4-domains-average · How It Works · Data refreshed daily, snapshot 2026-11-05.