Understand this story
On November 11, 2026, Tether AI Research released the on-device agent model BotPsy on Hugging Face. The model is post-trained from Qwen3.5-0.8B, with fewer than one billion parameters; BFCL v4 rose from 23.4 for the base model to 44.0, ACEBench from 30.1 to 58.7, and τ-bench from 23.2 to 51.5. Across six tool-calling and agent benchmarks, it surpasses the Qwen3.5-2B-Think, which has 2.5 times the parameter count, on five of them, trailing only on Claw-Eval (46.5 versus 48.9). The model underwent four-stage training consisting of SFT, multi-teacher online distillation, and two rounds of RL, and is released with open-source weights under Apache 2.0 in three precisions—bf16, Q8, and Q4—available in safetensors and GGUF formats. The next day, Unite.AI reported that the model was released under the name QVAC BotPsy-0.8B, with ToolSandbox rising from 30.5 to 64.6 and multi-turn BFCL v4 at 46.1; the model card also notes that its GPQA Diamond and MMLU-Pro scores are lower than those of the 2B model, and that unsafe responses accounted for 21% in automated red-team testing.
AI-generated from public reports
Latest development · Report
Tether Releases QVAC BotPsy-0.8B On-Device Tool-Calling ModelDevelopments
1 confirmed developments
Tether AI Research releases on-device assistant model BotPsy
Source:Hugging Face
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Nov 12
Unite.AI
Tether Releases QVAC BotPsy-0.8B On-Device Tool-Calling Model
Tether AI Research has released the open-weight model QVAC BotPsy-0.8B, built on Qwen3.5-0.8B and focused on function calling and multi-step agent workflows on personal devices, claiming it surpasses all tested models under 1.2B parameters on six agent benchmarks. BFCL v4 rose from the base model's 23.4 to 44.0, ACEBench from 30.1 to 58.7, ToolSandbox from 30.5 to 64.6, and multi-turn BFCL v4 reached 46.1. The GGUF version offers three files—BF16, Q8_0, and Q4_K_M—while the main repository releases BF16 safetensors weights under the Apache 2.0 license; the model card also notes that its GPQA Diamond and MMLU-Pro scores are lower than those of 2B models, and that unsafe responses accounted for 21% in automated red-team testing.
Nov 11
Hugging FaceFirst-party
Tether AI Releases BotPsy: A 0.8B On-device Agent Model Whose Tool-Calling Benchmarks Double Those of the Base Model
Tether AI Research has released BotPsy on Hugging Face, a sub-one-billion-parameter on-device agent model post-trained on Qwen3.5-0.8B, with BFCL v4 rising from 23.4 to 44.0, ACEBench from 30.1 to 58.7, and τ-bench from 23.2 to 51.5. Across six tool-calling and agent benchmarks, it surpasses Qwen3.5-2B-Think, which has 2.5 times the parameter count, on five of them, falling behind only on Claw-Eval (46.5 vs. 48.9). The model was trained in four stages: SFT, multi-teacher online distillation, and two rounds of RL, and will be released as open-source weights under Apache 2.0 in three precisions—bf16, Q8, and Q4—available in safetensors and GGUF formats.
Heat trend for this story
- Current
- 7
- Period peak
- 14 Nov 12 15:00
- 24-hour change
- —
Move the pointer or select the chart to inspect hourly heat. Use the left and right arrow keys with a keyboard.