Tether releases QVAC BotPsy-0.8B, an open-weight on-device agent

BotPsy is a post-trained version of Alibaba's Qwen3.5-0.8B thinking model. Tether describes it as built on the open Qwen3.5 architecture, and the model card names Qwen3.5-0.8B (Thinking) as the base[1][2]. It is text only, trained and tested in thinking mode, and English is the recommended language. Tether says the size suits local deployment on tasks of low to medium difficulty[1][2].

The card says Tether Data licenses the weights under the Apache 2.0 licence. The model was developed to advance the state of the art and to support research and educational uses, and it is not for safety-critical decisions or for uses involving money, personal information, health and other sensitive topics[2]. Training code and data are not included. Post-training was four stages: supervised fine-tuning, multi-teacher on-policy distillation, and two rounds of reinforcement learning[2].

BF16 safetensors are on Hugging Face. A second repository holds GGUF files for llama.cpp and the QVAC SDK: about 1.52 GB in unquantised BF16, about 0.81 GB in Q8_0, which Tether recommends, and about 0.53 GB in Q4_K_M[2][3].

Every figure below is from Tether's own testing, not an independent leaderboard. The announcement covers six agent benchmarks: BFCL v4, ACEBench, tau-bench, Claw-Eval, BrowseComp-Plus and ToolSandbox. On those six, Tether says BotPsy beat every other model it tested at 1.2 billion parameters or fewer[1][2].

On four of them, Tether says BotPsy roughly doubles its base, Qwen3.5-0.8B-Think, and beats Qwen3.5-2B-Think, a model 2.5 times its size[1]. The card prints the larger model's scores as well[2].

• BFCL v4: 23.4 for the base, 44.0 for BotPsy, 37.7 for Qwen3.5-2B-Think.
• ACEBench: 30.1, 58.7 and 55.4.
• tau-bench: 23.2, 51.5 and 46.5. The card defines that overall figure as the mean of retail, telecom and banking, excluding airline[2].
• ToolSandbox: 30.5, 64.6 and 41.2.

On the other two, the pattern breaks. BrowseComp-Plus, capped at 30 steps, is 4.5 for BotPsy, 2.8 for Qwen3.5-2B-Think and 1.9 for the base: the best of the five models Tether ran, and still a low score. Claw-Eval goes to Qwen3.5-2B-Think, 48.9 to 46.5[2].

The gap Tether highlights is multi-turn tool use on BFCL. BotPsy scores 46.1, against 23.8 for Qwen3.5-2B-Think and 7.5 for the base[1]. On the telecom slice of tau-bench, Tether reports 82.5, against 69.3 for Qwen3.5-2B-Think and 74.0 for MiniCPM5-1B-Think[1]. On BFCL v4's hallucination tests, which ask whether a model declines a tool call when none fits, BotPsy scores 86.1, the highest Tether tested[1]. The card says all five models ran in thinking mode on one harness, and that benchmarks calling an external API were run on 3 November 2026[2].

BotPsy joins the QVAC family beside VisionPsy for vision and language, TranscribePsy for speech, MedPsy for health-related tasks, and TranslatePsy for translation[1]. The longer-term sketch is a phone as the everyday companion, with a home PC or trusted peer-to-peer nodes taking heavier work, such as reading a document or organising a trip, and sending the actions back. Searches would go out over encrypted peer-to-peer links[1].

Paolo Ardoino, Tether's chief executive, said an assistant for a calendar, messages and documents needs personal information. "Running AI locally can reduce the need to send that information to centralized providers, giving people greater control over how it is processed." He said users should decide what the assistant can reach, and that developers should be able to run and adapt a model without depending on one company's servers[1].

Most agent products still assume a very large model in someone else's data centre. BotPsy is the other bet: a small model specialised for tool calls, published as GGUF files for a phone or a laptop, and scored by its maker against models up to 2.5 times its size. The scores are Tether's. The four core tests move a long way from the base model, and Claw-Eval still goes to the larger Qwen model. Tether issues the USDT stablecoin, and this model does not move money. The card positions the weights for research and education, and says they are not for uses involving money, personal information or health. In this release, a stablecoin company is publishing a local agent.