Tether’s 0.8B-Parameter BotPsy Beats a Model 2.5 Times Its Size at Tool Use, Making the Case for AI Agents That Run on Your Phone Instead of the Cloud

The race in AI agents has mostly been about bigger models running in bigger data centers. A new release from Tether’s AI research group points the other way. On November 11, the company behind the USDT stablecoin published BotPsy, a 0.8-billion-parameter language model trained specifically to call tools and complete multi-step tasks, small enough to run on a phone or laptop with no cloud connection at all.

The headline claim is that this tiny model beats every model of up to 1.2 billion parameters that Tether tested on six tool-use and agent benchmarks, and tops a sibling model two and a half times its size on five of them. If those numbers hold up under outside testing, they strengthen a business case many companies are already weighing: not every AI task needs a frontier model, and some of the most repetitive ones may not need the cloud.

What Tether Actually Released

BotPsy is built on top of Alibaba’s open Qwen3.5-0.8B model and further trained for function calling and agent workflows. According to Tether’s Hugging Face post, it is meant to act as a personal assistant that uses tools for simple, everyday jobs across areas such as web search, travel, commerce and finance, customer operations, device control, documents, calendars, email and messaging.

Tether says the weights are being released under the Apache 2.0 license for research and educational use, in three precisions (bf16, Q8 and Q4), as Hugging Face safetensors and as GGUF files that work with llama.cpp and Tether’s own QVAC SDK. The company is explicit about the limits: the model is best for low-to-medium difficulty tasks, was trained and tested mainly in English, scores much lower in Hindi and Russian, and is not intended for high-stakes uses involving money, personal information or health.

The Numbers Behind the Claim

All figures below are Tether’s own, published with its evaluation footnotes, and have not yet been independently reproduced.

• Roughly double its base model: On the Berkeley Function Calling Leaderboard (BFCL v4), the score rose from 23.4 to 44.0. On ACEBench it went from 30.1 to 58.7, and on τ-bench (retail, telecom and banking) from 23.2 to 51.5.
• Beating a bigger sibling: Against Qwen3.5-2B-Think, a model 2.5 times larger, BotPsy led on BFCL v4 (44.0 vs 37.7), ACEBench (58.7 vs 55.4), τ-bench (51.5 vs 46.5), ToolSandbox (64.6 vs 41.2) and BrowseComp-Plus (4.5 vs 2.8). The larger model won only on Claw-Eval (48.9 vs 46.5).
• Customer-service conversations: On the telecom portion of τ²-bench, where an agent has to hold a long customer conversation and follow a strict policy, BotPsy scored 82.5 compared with 69.3 for Qwen3.5-2B-Think.
• Knowing when not to act: On BFCL v4’s hallucination tests, which check whether a model declines when no suitable tool exists, it scored 86.1, the highest of the models tested.
• Less wasted thinking: On questions all models answered correctly, it used about three times fewer reasoning tokens than its base on GPQA Diamond (3,690 vs 11,563), and its rate of getting stuck in repetitive reasoning loops on IFBench fell from 22.9% to 1.4%.

Safety also improved by Tether’s measure. In automated red-teaming across six harm categories, BotPsy gave unsafe responses 21% of the time versus 36% for its base model. Under prompt-injection attacks, the type of attack that matters most for agents reading untrusted web pages or tool outputs, the unsafe rate dropped from 46% to 31%. That is better, but it still means nearly one in three injection attempts succeeded in testing, which is a strong reason to keep a model this size away from sensitive actions.

Why a Stablecoin Company Is Building Tiny AI

Tether frames the release as a response to the economics and risks of depending on cloud AI. Its post argues that frontier cloud models carry high and unpredictable operating costs, that providers can change, restrict or retire access at any time, and that sending data to third parties raises privacy and control concerns. Small, specialized models running locally, it argues, can handle many routine tasks at a fraction of the cost.

Tether is not alone in chasing cheaper agent “workers.” Two days earlier, Meta launched Muse Ember, its smallest cloud model, aimed at the repetitive work agents do all day, priced at $0.10 per million input tokens and $0.50 per million output tokens. The two releases take different routes, a cheap hosted API versus free weights on your own hardware, but they share the same premise: a large planner model does the hard thinking, and a small, inexpensive model handles the volume.

What It Means for Business

For companies building AI into products and operations, the takeaway is less about BotPsy itself and more about the direction of travel.

Cost per task is becoming the metric that matters. A model that reaches the same answer with a third of the reasoning tokens, and that rarely loops, is cheaper and faster everywhere it runs. On a device, that also means less battery drain and lower latency.

Privacy-sensitive workflows get a new option. Calendar, email and file tasks that a company is reluctant to send to an outside provider could, in principle, run entirely on an employee’s device. Teams in regulated industries should still note Tether’s own warning against using the model for personal-information or financial decisions.

Vendor lock-in becomes easier to hedge. Open weights under a permissive license mean a business can keep a fallback model it controls, even if a cloud provider changes prices or terms. The research-and-education framing of this release means legal teams should read the license terms carefully before any commercial deployment.

Benchmarks are not production. Every number in the release is self-reported, and the model’s best scores still leave plenty of room for error. A sensible approach is to test small models on narrow, low-risk tasks first, such as routing requests, filling forms or summarizing documents, and keep a larger model or a human in the loop for anything that touches money or customer data.

The Bottom Line

BotPsy shows how much agent capability can now be squeezed into a model under one billion parameters. It will not replace frontier models, and Tether does not claim it should. But together with cheap hosted options like Meta’s Muse Ember, it signals that the next phase of AI adoption may be decided less by who has the biggest model and more by who can run the right-sized model for each job, at the lowest cost, closest to the user.

Sources: Tether AI Research on Hugging Face; Meta, Muse Ember announcement.

Edited by Coron @QUE.COM
Website: https://QUE.COM Intelligence
Sponsored by: https://MAJ.COM AI Autonomous