Tether released QVAC BotPsy, an 800-million-parameter open-weight model built to act on a phone: change settings, send messages, manage a schedule, and recover when a step fails. Tether says it beat larger models on tool-use benchmarks. That is a claim, not a finding, and every number below comes from Tether's own documentation as reported by Decrypt. The take: a small agent that runs locally is a real shift for builders. A leaderboard row from the vendor is not the evidence for it.
What Tether says it shipped
According to Decrypt, BotPsy is built on Alibaba's Qwen3.5-0.8B architecture and trained further for tool use. It is meant to identify which tools a task needs, run commands in order, and adjust when something breaks.
The example Tether offers: an assistant has to send a message, but Wi-Fi is off and the inbox is full. BotPsy turns on Wi-Fi, clears storage, and finishes the job. The base model keeps retrying the same failed operation. It is a single described scenario, not a broad real-world test.
The numbers, with their provenance
Tether's technical documentation reports these scores:
- ▸BFCL v4: BotPsy 44.0%, Qwen3.5-0.8B base 23.4%, Qwen3.5-2B 37.7%.
- ▸ACEBench: BotPsy 58.7%, Qwen3.5-0.8B base 30.1%, Qwen3.5-2B 55.4%.
- ▸Across the set: five models, six tool-use evaluations, BotPsy highest on five. Its average was 45.0% against 38.8% for the larger Qwen model.
The "2.5 times its size" in the headline matches the ratio of the 2B model to BotPsy's 0.8B. Decrypt does not name the other models Tether tested. It also doesn't make clear which model the 38.8% average belongs to, and it describes scores only as a weighted percentage of successful performance. The article cites no independent evaluation.
So the honest reading is: Tether reports that a tuned 0.8B model scores above an untuned 2B sibling on two public tool-calling benchmarks. That is plausible. Targeted tuning often moves a small model a long way on the exact task it was tuned for. It is also unverified, and a benchmark is not the same thing as a phone in someone's pocket.
Why a builder should care anyway
Most agent stacks assume a cloud model on the other end of an API. Tether says BotPsy can handle multi-step phone tasks potentially without sending personal data to cloud servers. If that holds up, it changes the architecture options:
- ▸Local first, cloud for the hard parts. A small on-device model handles routine tool calls. A bigger model gets only the steps that need it.
- ▸Smaller memory footprint. Tether says it released compressed versions to cut memory needs for local deployment.
- ▸Open weights under Apache 2.0. Developers can download, modify and integrate it. Per the model card, training code and training data are not included, so this is open weights, not a reproducible training run.
The same distinction applies to any agent that touches money. A model that decides which tool to call is one layer. The wallet controls, spend limits and payment rails around it are another, and a benchmark score says nothing about them. If an agent can move funds, "it scored well on BFCL" is not a permissions model.
Where this sits in Tether's QVAC push
Decrypt places BotPsy inside QVAC, Tether's on-device AI effort, introduced in May 2025. The article lists earlier pieces: a 41-billion-token dataset and QVAC Workbench in October 2025, an SDK for offline AI apps on mobile and desktop earlier in 2026, and QVAC MedPsy, a 1.7B medical model that Tether says beat Google's MedGemma-27B on HealthBench Hard. Longer-term goals reportedly include agents that make payments in Bitcoin and USDT. That is a stated direction, not a shipped feature. No release date has been announced for a consumer assistant built on BotPsy.
What to watch
- ▸Independent evals. Does anyone outside Tether reproduce the BFCL v4 and ACEBench numbers, and name the full model lineup?
- ▸Real device runs. Failure recovery on actual phones, across messy settings, not one scripted Wi-Fi scenario.
- ▸The payments step. Whether QVAC agents ever sign transactions, and what limits sit between a model's output and a signature.
- ▸Tool-calling in your own stack. If you are building agents, the model is one pick among many. Sato Hub maps the skills, MCPs, wallets and payment rails around it, so you can compare the pieces on evidence, not on a launch post.
Build the agent you actually need at satohub.ai/build.