tether botpsy agent: 0.8B On-Device Model Beats a 2B Rival

The tether botpsy agent is a tiny open-weight model that Tether's AI unit released this week to run tool-using tasks right on your device. At 800 million parameters it is microscopic next to frontier chatbots, yet Tether's own benchmark table shows it beating a model 2.5 times its size on five of six agent tests. The catch is in the fine print: Tether says it is not meant for anything involving money.

What actually happened

Tether, the company behind USDT, has been quietly building an AI stack called QVAC that runs models on hardware users control instead of a cloud. Decrypt reported on November 11 that the newest piece, BotPsy, is an agentic model: it is trained to pick the right tool, call it with the right arguments, and recover when a step fails.

The weights landed on Hugging Face under the qvac organization, along with quantized GGUF files for llama.cpp and Tether's own QVAC SDK. The BotPsy model card describes it as a "personal assistant" for simple tasks in areas like calendar, messaging, files, device settings, travel and shopping. It thinks before it answers and works best in English.

The example Tether is pushing

Decrypt described one benchmark scenario Tether highlighted. An assistant had to send a message, but Wi-Fi was off and the inbox was full. BotPsy switched Wi-Fi on, cleared space and finished the job, while its base model kept retrying the same failed step. Boring? Yes. Also exactly the kind of chore that makes a phone agent useful.

The numbers, from Tether's own table

All five models in the comparison were tested in thinking mode with one shared harness, according to the card. Highlights:

• BFCL v4: BotPsy 44.0, Qwen3.5-0.8B base 23.4, Qwen3.5-2B 37.7
• ACEBench: 58.7 vs 30.1 vs 55.4
• τ-bench (average of retail, telecom, banking): 51.5 vs 23.2 vs 46.5
• ToolSandbox: 64.6 vs 30.5 vs 41.2
• Claw-Eval: the only loss, 46.5 vs 48.9 for Qwen3.5-2B

The card also reports strong instruction following, with IFEval rising from 59.7 to 90.8 over the base, and an 86.1 score on BFCL's hallucination tests, which check that a model declines when no suitable tool exists. That last one matters for agents: knowing when not to act is half the job.

These are vendor-run benchmarks. They are useful for comparing small models on the same harness, but they are not independent verification.

What is missing

Decrypt noted that the training code and training data are not included in the release, and Tether has not announced a date for a finished consumer assistant built on BotPsy. The card also says the model scores much lower in Hindi and Russian than in other tested languages.

Why crypto people care

QVAC's pitch since 2025 has included agents that can eventually interact with apps and even pay with Bitcoin and USDT. That is the "agents with wallets" dream half of crypto Twitter keeps shilling. But Tether's own card says BotPsy "is not intended for safety-critical or high-stakes automated decisions, or for applications involving money." So the tether botpsy agent is a step toward on-device agents, not a wallet bot you should trust with funds today.

It also fits a broader trend of small, local agents. We saw a similar angle with the mika android agent, and the risk side of letting agents act for you came up in apple agent risks.

What to do as a reader

If you want to play with the tether botpsy agent, keep it in sandbox mode:

1. Run it locally using the GGUF builds if you want privacy; that is the point of the project.
2. Give it low-stakes tools first: calendar, notes, device settings. Do not wire it to exchange keys or a hot wallet.
3. Check results yourself. A 0.8B model makes mistakes, and the card says so plainly.
4. Watch for independent evals. Vendor tables are a starting point, not the final word.

For background on the company funding all this, see our tether q3 profit breakdown. Not financial advice, just a nudge to keep your keys away from experimental bots.