Tether built an AI model that fits on your phone and still beats a bigger one
Tether, the company behind the USDT stablecoin, has released BotPsy, a 0.8-billion-parameter model meant to run personal-assistant tasks on your own device instead of in the cloud, and its benchmark charts put it ahead of a model 2.5 times its size.
Tether makes its money from a stablecoin. On November 11, the company’s Tether AI Research unit published a language model small enough to run on a phone, built on Alibaba’s Qwen3.5-0.8B, and the benchmark numbers it published are the kind that usually come from a much bigger lab.
The model is called BotPsy, built by a unit the company calls Tether AI Research. It’s tiny by current standards: 0.8 billion parameters, built on top of Alibaba’s Qwen3.5-0.8B. Tether’s own blog post, published November 11, says BotPsy is meant for “function calling and agentic workflows.” That means it’s not a chatbot you argue with about poetry. It’s a model meant to call tools: search the web, book travel, move money, open a file, read your calendar, send a message, all without the request leaving your device.
My read: a stablecoin issuer building a privacy-pitched on-device assistant model is a strange pairing on paper, but it tracks with what Tether’s been doing under the QVAC brand, which is trying to build decentralized, local-first AI infrastructure rather than compete with OpenAI on frontier scale.
What the benchmarks actually say
According to Tether’s blog post, BotPsy roughly doubles its base model’s tool-use scores. BFCL v4, a benchmark for function-calling accuracy, rises from 23.4 for the base Qwen3.5-0.8B to 44.0 for BotPsy. ACEBench goes from 30.1 to 58.7. Tau-bench, which scores multi-turn agent conversations across retail, telecom and banking, goes from 23.2 to 51.5.
Tether says BotPsy beats every model it tested at 1.2 billion parameters or fewer on all six agentic benchmarks it ran: BFCL v4, ACEBench, tau-bench, ToolSandbox, Claw-Eval and BrowseComp-Plus. More striking, per the same post, BotPsy also beats Qwen3.5-2B-Think, a model 2.5 times its size, on five of those six.
| Benchmark | BotPsy (0.8B) | Qwen3.5-2B-Think (2B) |
|---|---|---|
| BFCL v4 | 44.0 | 37.7 |
| ACEBench | 58.7 | 55.4 |
| tau-bench | 51.5 | 46.5 |
| ToolSandbox | 64.6 | 41.2 |
| BrowseComp-Plus | 4.5 | 2.8 |
| Claw-Eval | 46.5 | 48.9 |
The one benchmark where the bigger model wins is Claw-Eval, 48.9 to 46.5. Everywhere else in that table, the smaller model leads.
My read: these are Tether’s own numbers, run on Tether’s own test setup. No outside lab had re-run them as of publication. That matters. The direction here, a well-tuned small model beating a lazier big one on narrow tasks, isn’t new; Hugging Face has published several similar “small model punches up” stories this year. What’s genuinely notable is the safety number. Tether says BotPsy produces unsafe responses in automated red-teaming 21% of the time, versus 36% for its own base model and 47% for MiniCPM5-1B-Think, a rival. Under prompt-injection attacks specifically, the kind that matter most for an agent reading untrusted web pages, the unsafe rate dropped from 46% to 31%. Still, that’s roughly one in three attacks getting through. Keep this model away from anything with real stakes attached. Don’t trust it with your bank login.
Where it actually fits
BotPsy is explicitly not for hard problems. Tether’s own writeup says it’s “best suited to tasks of low to medium difficulty” and isn’t meant for safety-critical or high-stakes use, health, money, or personal information handled without a human checking the output. It’s trained mainly in English, and while it was tested on seven other languages through a benchmark called Multi-IF, Tether says it scores much lower in Hindi and Russian.
The weights are already live on Hugging Face under the qvac organization (BotPsy-0.8B in BF16 safetensors, plus GGUF builds in Q8 and Q4 for running through llama.cpp or Tether’s own QVAC SDK), licensed Apache 2.0. That’s a real, usable release today, not a promise for later; the blog post’s phrasing (“will be released”) reads like it was written before the Hub upload went live.
Note on sources: BotPsy's benchmark comparisons come entirely from Tether's own published post and its Hugging Face model card, both of which I read in full. I haven't found an independent third-party benchmark run of BotPsy as of this writing, so treat the "beats a model 2.5x its size" framing as the vendor's claim until someone else reproduces it.
Why a small model matters more than it sounds
The pitch for shrinking models down isn’t really about saving money, even though Tether leads with that: cloud APIs can change price or access terms overnight, and a model running on your phone can’t be switched off by someone else’s outage. The more concrete case is latency and privacy for simple, repeated tasks, checking a calendar, parsing an email, setting a reminder, where sending the request to a cloud model and waiting for tokens to stream back is slower than just running it locally. Tether reports BotPsy reaches correct answers using about a third of the reasoning tokens its base model needed on GPQA Diamond (3,690 tokens versus 11,563), which, if it holds up outside Tether’s own tests, is the kind of efficiency gain that actually shows up as a snappier assistant rather than a chart.
It’s part of a broader pattern worth watching: small, specialized models for narrow jobs, paired with occasional calls to a bigger cloud model for anything genuinely hard. Underdog’s Woof text-to-speech model, which fits in 32MB and runs offline in a browser, is the same bet applied to a different kind of model.
What I’d watch next: whether an outside group reruns BotPsy’s benchmark claims, and whether Tether’s QVAC SDK actually ships it inside a real consumer product rather than leaving it as a Hugging Face download for developers.
Related: llama.cpp v1.0.0 and its new versioning rule, the runtime that BotPsy’s GGUF builds run on.
Sources