🔍 Read the full analysis: BotPsy: Building An On-Device Assistant With A Compact AI Model on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Tether AI Research has introduced BotPsy, a 0.8-billion-parameter model based on Qwen3.5-0.8B for tool use and agent workflows, including possible on-device deployment. The company reports gains on six tool-use benchmarks, but the results have not been independently verified and do not establish how BotPsy performs on everyday devices or tasks.
Tether AI Research has introduced BotPsy, a 0.8-billion-parameter model built on Qwen3.5-0.8B for function calling and agent workflows, with the company presenting it as suitable for on-device use. Tether reports higher scores than the base model on several tool-use benchmarks, but the figures are company evaluations and do not show how reliably BotPsy will work in everyday deployments.
BotPsy is intended for low- to medium-difficulty tasks, according to Tether AI Research. The company describes it as a personal assistant that can use tools for activities such as web research, travel and lifestyle requests, commerce, business operations, device controls, documents, calendars and messaging. The announcement says the model is not intended for high-stakes use and acknowledges that it can make mistakes.
In results reported by Tether, BotPsy scored 44.0 on BFCL v4, compared with 23.4 for its Qwen3.5-0.8B-Think base model. It scored 58.7 on ACEBench, versus 30.1, and 51.5 on τ-bench, versus 23.2. Tether says the τ-bench result is the mean across retail, telecom and banking subsets. It also reports that BotPsy led models with up to 1.2 billion parameters on six tested tool-use benchmarks.
Against the larger Qwen3.5-2B-Think, results were mixed: BotPsy scored higher on five of the six reported tests, including BFCL v4, ACEBench, τ-bench, ToolSandbox and BrowseComp-Plus. The larger model scored 48.9 on Claw-Eval, compared with BotPsy’s 46.5. The supplied material also reports instruction-following and some knowledge-benchmark gains, while identifying weaker performance as instructions accumulate and in lower-resource languages.
BotPsy brings tool use to a compact model
Tether AI Research says its small assistant is designed for routine agent workflows and on-device use. Its benchmark results are developer-reported and do not establish everyday reliability.
Built for routine assistant work
The model is tuned for function calling and agent workflows, rather than frontier-scale general performance.
Find and gather information
Example uses include web research and travel requests.
Handle everyday workflows
It is presented for commerce, business operations, documents, calendars, and messaging.
Run on-device
Local deployment could reduce reliance on cloud services for routine tasks.
Reported results improved over its base
Tether AI Research reports higher scores for BotPsy on three named benchmarks.
| Benchmark | BotPsy | Qwen3.5-0.8B-Think |
|---|---|---|
| BFCL v4 | 44.0 | 23.4 |
| ACEBench | 58.7 | 30.1 |
| τ-bench | 51.5 | 23.2 |
The larger-model comparison was mixed
BotPsy scored higher on five of six tool-use tests against Qwen3.5-2B-Think.
- BFCL v4
- ACEBench
- τ-bench
- ToolSandbox
- BrowseComp-Plus
- Qwen3.5-2B-Think scored 48.9 on Claw-Eval, compared with BotPsy’s 46.5.
Limits matter for deployment
The supplied material describes areas where performance may fall short and leaves practical questions open.
- Results come from the developer; independent reproduction is not provided.
- Performance drops as instructions accumulate and lags in lower-resource languages.
- Hardware needs, memory use, power use, and real-world reliability are not established.
- Tether says BotPsy can make mistakes and is not intended for high-stakes use.
Why Local Assistant Models Matter
BotPsy’s proposed value is that a relatively small model could handle some routine tool-using tasks on a user’s device, rather than sending every request to a cloud service. If it performs adequately in practice, local operation could reduce dependence on external providers for those requests and may keep some user data on-device. The announcement does not establish that those privacy or cost benefits occur in a particular deployment.
Tool use also creates risks beyond producing an incorrect answer: an assistant must select appropriate actions and recognize when its available tools cannot fulfill a request. Tether reports an 86.1 score on BFCL v4 hallucination tests, which it says assess abstention when no suitable tool exists. It also reports a 31% unsafe-response rate under prompt-injection attacks, down from 46% for the base model. These are vendor-reported test results, not proof of safety across devices, tools or user situations.
The announcement therefore matters as a development in compact agent models, but the practical case remains conditional. Benchmark performance can help compare systems under defined tests; it cannot by itself show that an assistant will complete a changing, multi-step task reliably or handle sensitive information safely.
Products Worth Considering
Amazon Kindle Scribe (3rd Generation), 64GB, 11-inch Display, Graphite, with Premium Pen
Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation
LG UltraGear 25G590B-B 25-Inch Full HD 1000Hz Gaming Monitor
As an affiliate, we earn on qualifying purchases.
BotPsy’s Model and Distribution
BotPsy is based on Qwen3.5-0.8B and is distributed as model weights, according to the source material. Listed formats include safetensors and GGUF, with bf16, Q8 and Q4 variants. Its stated focus is function calling and agent workflows rather than frontier-scale general-purpose performance.
Tether’s rationale is that specialized small models may handle routine work with lower operating costs and more local control over data than cloud systems. That is the developer’s argument, not a verified cost comparison. The supplied material provides selected benchmark scores but no independent evaluation or user deployment results. It also does not give a release date or describe ongoing support and update commitments.
“BotPsy is a “personal assistant” that uses tools to perform a variety of simple tasks for the user.”
— Tether AI Research
Open Questions on Real-World Use
The reported benchmark results have not been independently evaluated in the supplied material. It does not include full testing protocols, so it is unclear whether the results would be reproduced by other evaluators or across different implementations. Nor do the scores show how BotPsy handles changing tools, ambiguous requests, long workflows or sensitive user data.
On-device requirements are also unspecified. The announcement gives no measurements for memory use, power consumption or latency across representative phones and computers, and offers no comparison of user costs against cloud models. Although the source says performance weakens as instructions accumulate and in lower-resource languages, it does not quantify those limits. The reported prompt-injection rate likewise should not be read as a guarantee of protection in other settings.
Evidence Needed From Deployments
The next useful step would be independent reproduction of the benchmark results and deployment testing on representative devices. Such evaluations could report latency, memory and energy use, tool-call reliability, and performance when a workflow runs into errors or requires several steps.
People considering BotPsy for practical use will also need clear documentation on supported runtimes, safeguards and update practices. Tether cautions against high-stakes use, but the source material provides no timeline for further releases, testing or deployment reports. Until that evidence is available, the benchmark figures describe the company’s evaluations rather than established performance in users’ hands.
Key Questions
What is BotPsy?
BotPsy is a 0.8-billion-parameter model from Tether AI Research, based on Qwen3.5-0.8B and designed for function calling and agent workflows.
Can BotPsy run on a device?
Tether presents BotPsy for on-device deployment and lists model-weight formats including safetensors and GGUF. The supplied information does not specify device requirements or measured performance on phones or computers.
Are BotPsy’s benchmark results independently verified?
The scores in the supplied material are reported by Tether AI Research. It does not identify an independent evaluation or provide enough detail to establish how results would reproduce across implementations.
Is BotPsy intended for high-stakes tasks?
No. Tether says the model is aimed at low- to medium-difficulty tasks, cautions against high-stakes use and acknowledges that it can make mistakes.
When will BotPsy be available?
The supplied material does not specify a release date or provide a timeline for further evaluations and deployment information.
Primary source: Hugging Face · via ThorstenMeyerAI.com
