Tether AI Research released BotPsy on Nov. 112, a 0.8 billion-parameter language model the company says is designed for on-device personal-assistant tasks and tool use.
The announcement matters because a small model that can call tools more reliably could move some assistant work from cloud services to phones or laptops. But the available evidence comes from Tether’s own pages, with no independent test or reproduction in the retrieved sources.
In a QVAC explainer on Tether’s site2, the company said QVAC is its local-AI platform and framed BotPsy as a response to cloud cost, privacy and service-dependence concerns. That page says references to Tether AI Research mean Tether Data, S.A. de C.V., and says BotPsy starts from the Qwen3.5-0.8B base model and was trained to make small models more reliable at calling tools.
According to Tether’s Hugging Face post1, BotPsy is built on Qwen3.5-0.8B, targets function calling and multi-step agent workflows, and is suited to low- to medium-difficulty tasks rather than high-stakes use. The company said the model can work across web search, travel, commerce, business operations, files, calendars and messaging, while warning that it still makes mistakes.
Tether said BotPsy “roughly doubles” the agentic performance of its base model on several benchmarks. In figures reported in the Hugging Face post1, BFCL v4 rose to 44.0 from 23.4 for Qwen3.5-0.8B-Think, ACEBench to 58.7 from 30.1, and τ-bench to 51.5 from 23.2.
The same post published a six-benchmark table with an unweighted average of 45.0 for BotPsy, compared with 38.8 for Qwen3.5-2B-Think, 25.3 for LFM2.5-1.2B-Think and 22.7 for MiniCPM5-1B-Think under Tether’s setup. Tether also said BotPsy beat Qwen3.5-2B-Think on five of those six benchmarks, but its own table shows Qwen3.5-2B-Think leading on Claw-Eval, 48.9 to 46.51.
Those comparisons are limited to models Tether said it tested and to an evaluation setup whose footnotes were referenced but not fully available in the retrieved Hugging Face text. The company describes BotPsy as ahead of every model of up to 1.2B parameters that it tested, but that remains a vendor-supplied result.
Reported gains beyond tool use
Tether also reported gains in instruction following and some knowledge benchmarks. In the Hugging Face post1, IFEval rose to 90.8 from 59.7 and IFBench to 55.9 from 22.0, while GPQA Diamond rose to 38.1 from 33.2 and MMLU-Pro to 48.9 from 41.6, again relative to the base model on the company’s evaluation.
On reasoning efficiency, Tether said BotPsy used 3,690 reasoning tokens per correct answer on GPQA Diamond against 11,563 for the base model, and 1,912 against 5,105 on MMLU-Pro, on questions that all five compared models answered correctly. The Hugging Face post said shorter reasoning “should reduce latency” on-device, and a QVAC explainer2 separately said the model answers faster because it uses fewer reasoning tokens, but neither company page reviewed reports measured device latency.
Tether also published safety-related figures. The Hugging Face post1 said automated red-teaming produced unsafe responses 21% of the time for BotPsy versus 36% for the base model, and that the unsafe rate under prompt-injection attacks fell to 31% from 46%. Those figures are benchmark outputs under the developer’s conditions and do not establish behavior across deployments.
A QVAC explainer2 said BotPsy is aimed at developers rather than end users and recommended English as the primary language. That page said the model also scores well in Chinese, French, Spanish, Italian and Portuguese, but is weaker in Russian and much weaker in Hindi on the multilingual instruction test it described.
Availability differs across Tether pages
Tether’s two pages differ on availability. The Hugging Face post1 said BotPsy “will be released” with open weights under the Apache 2.0 license for research and educational use, in bf16, Q8 and Q4 as Hugging Face safetensors and GGUF files. The QVAC page2 said the weights are already free to download on Hugging Face under Apache 2.0 and said the model is available in two precision variants, BF16 and 8-bit, as standard weights and GGUF files. The company pages do not resolve the timing difference or the status of Q4.
Tether’s phrasing also pairs Apache 2.0 with “research and educational use.” The model-card license terms were not reviewed here, so the scope of permitted use beyond Tether’s wording is not established by the retrieved documents.
The company pages reviewed present BotPsy as suitable for phones and laptops, but they do not report measured latency, memory use or power draw on specific devices. The QVAC page’s “answers faster” claim is tied to token counts rather than device timings.
Tether said future work will expand local services integration, peer-to-peer search through Hypersearch, and delegation of heavier tasks from a phone to home PCs or trusted peer-to-peer nodes, according to the Hugging Face post1.