Tether introduces an 800-million-parameter AI assistant built for on-device tasks

Tether AI Research has introduced BotPsy, an 800-million-parameter AI model designed to perform tasks through software tools on local devices. Built from Qwen3.5-0.8B, it targets low- to medium-difficulty work rather than demanding general-purpose assignments. Tether reports that the model beat every tested rival on five of six tool-use benchmarks; open weights are still forthcoming.

BotPsy is designed for function calling: choosing a software tool and supplying the inputs needed to carry out a request. Tether’s proposed uses include calendar and email tasks, file handling, commerce, travel, and device control. Its small size is central to the company’s pitch for on-device deployment, not a claim that it can handle every assistant job.

Demonstrations first, specialist teachers next

Tether trained BotPsy in four stages, starting with examples of tool-use conversations. Those examples covered single calls, parallel calls and sequences of actions, alongside cases where the model should not use a tool. This first stage improved tool use but reduced general knowledge and instruction-following performance relative to the starting model.

The next stage used three larger Qwen3.5-4B models as specialist teachers for tool use, reasoning and instruction following. BotPsy generated its own responses and received feedback from the relevant teacher on each text token. Tether says this process restored knowledge and instruction-following scores above the base model while improving tool use further.

Two rounds of reinforcement learning followed, rewarding answers checked by rules and verifiers. Tool calls earned partial credit for matching the correct tool, input names and input values, rather than receiving only a pass-or-fail grade. The final round broadened training to math, code, knowledge, structured outputs and multi-step tool use in a simulated workplace.

Tool-use gains, not across-the-board intelligence

In Tether’s evaluation, BotPsy outperformed the tested models with up to 1.2 billion parameters on all six tool-use and agent benchmarks. It also beat Qwen3.5-2B-Think, a model 2.5 times its size, on five. The exception was Claw-Eval, where the larger model scored 48.9 against BotPsy’s 46.5.

Tether’s reported gains over the base model

BFCL v4

23.4→44.0

score out of 100

Tether reports that BotPsy improved on Qwen3.5-0.8B-Think from 23.4 to 44.0 on BFCL v4.

ToolSandbox

30.5→64.6

score out of 100

On ToolSandbox, Tether reports a rise from the base model’s 30.5 to BotPsy’s 64.6.

Those rankings do not mean every task is well within reach. BotPsy’s BrowseComp-Plus score was just 4.5 out of 100, despite leading the comparison. General knowledge also remained a size-dependent weakness: on GPQA Diamond and MMLU-Pro, Qwen3.5-2B-Think scored substantially higher. The stronger tool-use results did not erase that gap.

Local deployment still needs safety limits

Tether’s automated adversarial tests found unsafe responses in 21% of cases, down from 36% for the base model. Under prompt-injection attacks—attempts to redirect the agent through untrusted tool outputs—the unsafe rate fell from 46% to 31%. Tether says BotPsy still makes mistakes and is not intended for high-stakes use.

Efficiency improved in a narrower, measurable way. On GPQA Diamond questions that all five models answered correctly, BotPsy used about one-third as many reasoning tokens as its base model. Across IFBench attempts, repetitive reasoning loops fell from 22.9% to 1.4%. Tether expects shorter reasoning to reduce on-device latency; that is an expectation, not a measured device-speed result.

For developers, the distribution plan remains future-tense. Tether says it will release open weights under Apache 2.0, describing research and educational use. The planned packages include bf16, Q8 and Q4 versions in safetensors and GGUF formats, with GGUF files intended for the QVAC SDK and llama.cpp.