Tether AI Research releases BotPsy, a 0.8B parameter agentic model outperforming larger rivals

• Tether AI Research released BotPsy, a 0.8B parameter agentic language model based on Qwen3.5-0.8B, available under Apache 2.0 with safetensors and GGUF formats.
• BotPsy outperforms Qwen3.5-2B-Think (2.5x larger) on 5 of 6 tool-use and agentic benchmarks, including BFCL v4 (44.0 vs 37.7) and ToolSandbox (64.6 vs 41.2).
• The model reduces repetitive reasoning loops from 22.9% to 1.4% and uses up to 3x fewer tokens for correct answers on knowledge tasks compared to its base model.
• Designed for low to medium difficulty on-device tasks, BotPsy is not intended for high-stakes or safety-critical applications.

Technical Performance

• Leads in multi-turn workflows with 82.5 on τ²-bench telecom and 46.1 on multi-turn BFCL v4.
• Achieves 90.8 on IFEval and 55.9 on IFBench, surpassing larger competitors in instruction following.
• Knowledge scores remain lower than Qwen3.5-2B-Think (GPQA 38.1 vs 59.9), reflecting the model's design to rely on tools for unknown facts.

Methodology and Limitations

• Trained via four stages: SFT, Multi-teacher On-Policy Distillation, and two rounds of RL using DAPO.
• Weaknesses include open-ended search (BrowseComp-Plus 4.5) and low-resource languages like Hindi (24.5).
• Evaluation conducted in thinking mode with external API benchmarks run on November 3, 2026.