Tether's BotPsy AI model beats larger rivals in agent benchmarks

Tether AI Research released QVAC BotPsy on November 11, 2026, an open-weight AI model with 0.8 billion parameters that outperformed models up to 2.5 times its size on tool-use benchmarks built for AI agents, according to the company's published results.

The stablecoin issuer best known for USDT is now claiming state-of-the-art results in the sub-1B class for agentic tasks — AI work that goes beyond chat, covering function calling, form filling, email handling and document management.

What did BotPsy actually score?

Across six agentic evaluations, BotPsy posted a mean score of 45.0. It beat Qwen3.5-2B-Think, a 2-billion-parameter model, on five of the six tests.

The improvements over its own base model were steep:

• BFCL v4 (function calling): 44.0, up from 23.4 for the base model
• τ-bench: 51.5, compared with 23.2 for the base
• Hallucination resistance on BFCL v4: 86.1 — the top result among competitors under 1.2B parameters

How was it built?

BotPsy is built on Alibaba's Qwen3.5-0.8B base model, fine-tuned with a range of learning strategies to sharpen its agentic skills. Tether targets it at everyday digital chores: email function calling and document handling.

The release ships under the Apache 2.0 license, a permissive open-source license allowing broad use and modification. Tether also published the model in GGUF formats, a file type designed for running models on local hardware.

Tether positions BotPsy primarily for research and educational purposes, and its own guidance says the model suits low-to-medium difficulty tasks. It joins the QVAC Psy family, which already includes VisionPsy and MedPsy.

Why does a stablecoin issuer build AI?

BotPsy sits inside Tether's broader QVAC initiative, which aims at decentralized AI running entirely on local devices with no cloud dependency. Tether frames the release as a step toward personal assistants handling everyday tasks on hardware users control.

For the AI industry, the model is another data point in a quiet but meaningful shift: efficiency gains in small models. If a 0.8B model beats a 2B rival on most agentic tests, the hardware bar for useful on-device assistants drops.

The results need perspective, though. Six evaluations make a focused scorecard, not a full report card, and BotPsy did not win all six head-to-heads against Qwen3.5-2B-Think. Tether's own documentation limits it to lower-difficulty work, so builders should treat it as a strong option for narrow, well-defined agent tasks rather than a general replacement for larger systems.

The launch carries no token mechanics, so there's no new asset attached. What it does add is substance to a Tether strategy extending well beyond stablecoins.

What comes next?

With Apache 2.0 weights and GGUF builds already public, developers can run BotPsy locally today — and the real test will be whether the QVAC ecosystem produces models that hold up on harder agentic workloads as the local-AI bet scales.