AI BriefingBotPsy-0.8B Achieves Higher Tool-Use Performance Than 2B Models
·2026.11.12 12:30
Key point
Tether AI released BotPsy-0.8B, based on Qwen3.5-0.8B, demonstrating superior tool-use performance compared to 2B models.
Details
- Tether AI Research released BotPsy-0.8B, a 0.8B parameter tool-calling model based on Qwen3.5-0.8B, under the Apache 2.0 license
- Averaging 45.0 points across six tool-use benchmarks, it outperforms Qwen3.5-2B (38.8 points), recording significant performance improvements over the base model on BFCL v4 (44.0 points) and ToolSandbox (64.6 points)
- Enhanced tool-calling accuracy and multi-turn processing capabilities through a four-stage post-training process involving SFT, MOPD distillation, tool-use RL, and multi-domain RL
- Uses 1.4 to 3.1 times fewer inference tokens per correct answer than the base model, achieving the highest scores among compared models on IFEval (90.8 points) and IFBench (55.9 points)
- Scores 17 to 22 points lower than Qwen3.5-2B on general knowledge benchmarks such as GPQA Diamond (38.1 points) and MMLU-Pro (48.9 points), confirming its design as tool-centric rather than a knowledge model
- Compatible with Hugging Face, vLLM, and llama.cpp, with Q8_0 quantization (0.81GB) recommended. English is recommended, Korean is unevaluated, and use for sensitive topics or dangerous automated decision-making is prohibited
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
PyTorch ExecuTorch Adds WebGPU Backend for In-Browser GPU Inference