Tether AI Research, operating through its QVAC initiative, released technical specifications and open model weights on November 11, 2026, for BotPsy, an on-device personal assistant driven by an 800-million-parameter (0.8B) agentic language model engineered specifically for local tool calling, system interaction, and privacy-preserving execution on edge hardware.
Four-Stage Training Recipe and Distillation Architecture
BotPsy is built upon Alibaba’s open Qwen3.5-0.8B base architecture but undergoes an intensive four-stage post-training pipeline designed to maximize tool-calling precision without degrading underlying language comprehension. The authors observed that standard fine-tuning for tool calling often causes catastrophic forgetting in general reasoning and instruction following. To counteract this degradation, QVAC structured the post-training lifecycle across four distinct phases:
- Supervised Fine-Tuning (SFT): Full-parameter fine-tuning on diverse tool-use conversations, including simulated multi-turn agent trajectories, parallel function calls, and deliberate negative cases where the model must refrain from invoking tools when no valid utility exists.
- Multi-Teacher On-Policy Distillation (MOPD): The student model generates candidate responses while receiving per-token feedback measured by reverse Kullback-Leibler (KL) divergence from three distinct Qwen3.5-4B specialist teacher models. These teachers were independently trained via reinforcement learning for tool use, general reasoning, and instruction following, restoring capabilities lost during SFT.
- Tool-Calling Reinforcement Learning: Using the DAPO reinforcement learning framework, the model is optimized with fine-grained, partial-credit reward verifiers patterned after ToolRL. Rather than applying binary success metrics, rewards evaluate predicted tool names, parameter keys, and argument values against ground-truth executions.
- General-Domain Verifiable Task RL: A concluding reinforcement learning phase trains the model across verifiable mathematics, code generation, multiple-choice STEM datasets, and simulated workplace tasks to ensure multi-step stability.
Benchmark Results and Efficiency Across Agentic Tasks
In benchmark evaluations reported by the authors, BotPsy-0.8B demonstrated substantial performance increases over its base model and exceeded the scores of larger models within the lightweight category. On the Berkeley Function-Calling Leaderboard (BFCL v4), BotPsy achieved an overall score of 44.0, up from 23.4 on the base Qwen3.5-0.8B-Think and exceeding the 37.7 score posted by Qwen3.5-2B-Think (a model with 2.5 times more parameters). On ACEBench function-calling evaluations, BotPsy reached 58.7 compared to 30.1 for the base model and 55.4 for Qwen3.5-2B-Think. Similar leads were reported on τ-bench (51.5 versus 23.2 base) and ToolSandbox (64.6 versus 30.5 base).
The developers emphasized that an effective autonomous agent must recognize when not to execute actions. On BFCL hallucination benchmarks, which test whether a model refrains from calling tools when no appropriate option is provided, BotPsy scored 86.1, the highest across tested models. In multi-turn customer troubleshooting on τ²-bench telecom tasks, the model achieved 82.5, surpassing Qwen3.5-2B-Think (69.3) while requiring 26% fewer assistant turns to resolve tasks.
Reasoning efficiency was also evaluated across academic benchmarks. BotPsy reached correct solutions while using 1.4 to 3.1 times fewer reasoning tokens than its base model on GPQA Diamond (3,690 versus 11,563 tokens) and MMLU-Pro (1,912 versus 5,105 tokens). On the IFBench instruction-following evaluation, the model’s propensity to enter repetitive reasoning loops dropped from 22.9% to 1.4%, and instances of hitting the 32,000-token context ceiling fell from 28.3% to 0.2%. Automated red-teaming across six harm categories demonstrated an unsafe response rate of 21% (down from 36% in the base model), while vulnerability to prompt-injection attacks decreased from 46% to 31%.
Stated Limitations and Weaknesses
The authors explicitly outlined several operational boundaries and performance limitations. BotPsy showed pronounced weaknesses in open-ended web research, recording a BrowseComp-Plus score of 4.5 and low scores on BFCL web search (14.0) and memory retention (16.3). Multi-turn instruction retention degraded noticeably over extended conversations: while turn-one performance on Multi-IF began at 84.3, it dropped to 51.0 by turn three. Furthermore, linguistic performance exhibited significant disparity; while English, Chinese, French, Spanish, Italian, and Portuguese scored between 72 and 80, performance in lower-resource languages dropped steeply, with Russian scoring 57.6 and Hindi reaching only 24.5. The developers noted that the model is designed for low-to-medium difficulty routines and is not intended for safety-critical, high-stakes decision-making, or unmonitored financial and healthcare tasks.
Implications for Academic Labs and Edge Research
The release of sub-billion parameter agentic architectures carries immediate significance for university research groups and higher education laboratories. In field research and clinical observational studies, running local agents directly on portable laptops or mobile hardware eliminates the compliance risk of transmitting participant audio, notes, or sensor telemetry to external cloud providers. This local containment directly supports compliance with institutional data management plans and institutional review board (IRB) privacy mandates.
From an infrastructure perspective, academic institutions facing escalating cloud API expenditures can deploy quantized sub-billion models locally to handle routine data transformation and local tool execution. Aligning departmental workflows with token cost economics in institutional planning enables departments to reserve frontier models for complex conceptual reasoning while delegating repetitive functional tasks to edge devices. Furthermore, open model releases enrich the ecosystem of AI and machine learning research outputs, providing verifiable testbeds that inform emerging agentic AI benchmark governance and institutional frontier AI safety governance frameworks.
Availability and Licensing
The BotPsy model weights have been released under the open-source Apache 2.0 license. The model artifacts are available on Hugging Face in safetensors format alongside GGUF quantizations (bf16, Q8, and Q4) compatible with llama.cpp and the QVAC software development kit, allowing researchers to run local tool-use agents on commodity consumer hardware.
Sources
- Tether AI Research (QVAC). “BotPsy: On-Device Personal Assistant with a Sub-Billion-Parameter Agentic Language Model.” Hugging Face Blog. Published November 11, 2026. Accessed 2026-11-14. https://huggingface.co/blog/qvac/botpsy
- Qwen Team. “Qwen2.5 / Qwen3 Technical Reports and Architecture Specifications.” Alibaba Group. Accessed 2026-11-14. https://huggingface.co/Qwen







