Tether's QVAC BotPsy Beats Larger AI Models in Tool-Calling Tests

Tether AI Research released QVAC BotPsy-0.8B on November 11, 2026, and reported that the 0.8-billion-parameter open-weight model outperformed every tested model up to 1.2 billion parameters across six agentic benchmarks. The model specializes in tool calling — connecting AI assistants to calendars, messaging, files, device settings, web research, travel and commerce — and it is built for local deployment on personal devices. The base is Alibaba's Qwen3.5-0.8B, and Tether positions the model as the action layer for a future personal assistant.

What do the benchmark numbers show?

Tether's model-card evaluation reports large specialization gains. BFCL v4 rose from 23.4 for the Qwen3.5-0.8B-Think base to 44.0 for BotPsy. ACEBench climbed from 30.1 to 58.7. ToolSandbox jumped from 30.5 to 64.6. All scores are percentages, and Tether evaluated all five models in thinking mode with a shared harness.

The largest gain concerns continuity. BotPsy scored 46.1 on BFCL v4 multi-turn, against 7.5 for its base and 23.8 for Qwen3.5-2B-Think. That test measures whether an agent keeps working as new information arrives — checking availability, handling a conflict, updating a calendar. BotPsy also scored 86.1 on BFCL's hallucination test, which measures whether the model refrains from a tool call when no available tool fits.

BotPsy beat the larger Qwen3.5-2B-Think on five of six agentic benchmarks. The 2B model led only on Claw-Eval, 48.9 to 46.5.

Where does the model fall short?

Tether's wider results do not establish general superiority. On GPQA Diamond, BotPsy scored 38.1 against the 2B model's 59.9. On MMLU-Pro, it scored 48.9 against 65.5. Its BrowseComp-Plus score of 4.5 led the five-model group but remains low in absolute terms, exposing weak open-ended research performance. BotPsy is designed to call tools for facts it does not know — a workaround, not a fix.

The safety data deserves more attention than the launch materials give it. In Tether's automated red-teaming across six harm categories, BotPsy produced unsafe responses 21% of the time, versus 36% for its base. Under prompt-injection attacks — a particular concern for agents reading untrusted tool outputs — the unsafe rates were 31% and 46%, respectively. The smaller model improved on both tests while still generating unsafe responses roughly one time in three under injection.

Tether's stated boundaries match that reality. BotPsy was trained and tested primarily in English. Its Multi-IF evaluation covered seven additional languages and produced much weaker instruction-following results in Hindi and Russian. The model card states it is not intended for safety-critical or high-stakes automated decisions, or for applications involving money, personal information, health and other sensitive topics.

How was it trained, and what ships today?

A technical write-up describes four post-training stages: supervised fine-tuning, multi-teacher on-policy distillation, and two reinforcement-learning rounds with rule-based and verifiable rewards. Three specialist Qwen3.5-4B teachers covered tool use, general reasoning and instruction following. The release omits the training code and training data — a notable gap for an open-weight project.

The model also reasons more economically. On GPQA Diamond questions all five models answered correctly, Tether measured 3,690 reasoning tokens for BotPsy versus 11,563 for its base; on MMLU-Pro, 1,912 versus 5,105. That is roughly a threefold reduction, useful for local inference.

The GGUF release supplies three files for llama.cpp and the QVAC SDK: BF16 at about 1.52 GB, Q8_0 at about 0.81 GB and Q4_K_M at about 0.53 GB. Tether recommends Q8_0, which averaged 50.7 across ten benchmarks — slightly above BF16's 50.4 at about half the file size. Q4_K_M averaged 48.9 and targets tighter memory budgets. BF16 safetensors weights ship under Apache 2.0.

What comes next?

BotPsy joins QVAC's vision, speech, health and translation models, with future development intended to combine their capabilities. "Users should decide what their assistant can access and which external services it can connect to," said Tether CEO Paolo Ardoino. Tether's longer-term design routes everyday interaction through an Android or iPhone, hands demanding tasks to a home PC or trusted peer-to-peer nodes, and sends web queries through encrypted peer-to-peer connections. That system remains a stated vision; the current release supplies the model and deployment files. Whether BotPsy becomes a useful assistant depends on application-level permissions, execution controls and integrations that do not yet exist.