Tether launches 800M-parameter QVAC BotPsy for on-device use

By Elena Vasquez

Tether, the issuer of USDT, launched QVAC BotPsy on Nov. 11 and reported that the 800-million-parameter model scored “nearly twice as high” as its base model in internal testing. Built on Alibaba’s Qwen3.5 architecture, QVAC BotPsy runs on users’ devices and is distributed with open weights.

The release gives developers access to model files rather than requiring reliance solely on a hosted interface. It also extends the USDT issuer into on-device artificial intelligence software, although Tether did not identify supported hardware, memory requirements, installation channels or usage terms.

Open weights do not by themselves resolve those questions, and the announcement did not say whether it made training code or complete reproducibility materials available.

What do the benchmark claims establish?

Tether said QVAC BotPsy scored “nearly twice as high” as Qwen3.5-0.8B-Think, its base model, across four major tool-use benchmarks. BFCL v4 and ACEBench were two of those benchmarks. Tether also said the model “outperformed Qwen3.5-2B-Think in those tests,” even though that comparator had 2.5 times as many parameters, or about 2 billion against 800 million.

That is a parameter-efficiency claim, not evidence of universal superiority. The announcement did not publish raw scores, benchmark settings, latency, run counts, statistical variance or independent replication. Without those details, outside researchers cannot determine how much of the reported gap comes from the model and how much from the evaluation setup.

Takeaway: The comparison is material, but it remains a company-run test without published evidence that independent testing would reproduce it.

What changes when the AI runs on-device?

Tether said QVAC BotPsy can reduce the need to send personal data, including emails and calendar information, to centralized service providers. Local processing can narrow that exposure, but the company did not say that all information remains on the device in every deployment. Its announcement also did not describe telemetry, data retention, prompt routing or security controls, so the privacy claim establishes a possible reduction rather than a guarantee.

The model also targets errors that produce calls to “nonexistent tools when no suitable tool is available for a request.” In tool-using systems, an invalid function call can interrupt an otherwise valid workflow. Tether disclosed no before-and-after error rates, however, leaving the size of the improvement unquantified.

Takeaway: On-device execution may reduce data exposure and failed tool calls, but the announcement proves neither outcome across real deployments.

What does the launch mean for Tether?

QVAC BotPsy gives Tether a product beyond USDT and introduces an open-weight distribution model for an on-device assistant. The announcement did not describe pricing, revenue sharing, commercial partnerships, an API, an adoption target or a deployment timetable. It also did not connect QVAC BotPsy to USDT transactions, payments or token utility.

That separation matters for readers assessing the release as a technology product rather than a crypto promotion. The announcement presents efficiency, privacy and reliability as design goals, not documented guarantees, and offers no independent validator or customer deployment evidence. Tether may seek a broader technology identity, but the announcement does not establish a commercial strategy.

Takeaway: QVAC BotPsy broadens Tether’s software activity without establishing a commercial link between the model and USDT.

What market change comes next?

The relevant market-structure test is whether compact, open-weight models can displace routine inference work now handled by centralized providers. QVAC BotPsy enters that contest with a small parameter count and local execution, but its commercial weight will depend on independent benchmark replication and performance on specified consumer devices. Tether set no validation deadline, making external testing the next credible checkpoint rather than another company-run score.

If independent results confirm the tool-use efficiency and local-data advantages, the practical shift would be from sending more tasks to remote services toward running more artificial intelligence locally. That would make on-device models a larger part of the AI market rather than a niche feature.

Takeaway: The next market-structure checkpoint is independent replication, because independently verified efficiency gains could move routine AI work from centralized providers onto user devices.