Skip to content

Tether launches BotPsy: lightweight device agent outperforms larger models

Share
Tether launches BotPsy: lightweight device agent outperforms larger models

Listen to this article

Read by Anchor

Rising inference costs and privacy risks in massive cloud models are pushing the AI sector toward local hardware solutions. The research unit at Tether Data announced the launch of the BotPsy model, a ultra-small language model of just 800 million parameters (0.8 B) built on the Qwen3.5-0.8B base and intended for function-calling tasks and directing smart agents directly on personal devices.

Calling efficiency outperforms a doubling of size.Standard benchmark results show BotPsy surpassing all competing models in the sub-1.2 B-parameter category, even exceeding the Qwen3.5-2B-Think model, which is two-and-a-half times larger, in five of six standard agent benchmarks. The model scored 44.0 points on BFCL v4 versus 23.4 for its base model and 37.7 for the larger model, achieved 58.7 on ACEBench, 51.5 on tau-bench, and 64.6 on ToolSandbox, while the larger model only led on Claw-Eval with 48.9 versus 46.5 points.

Training followed a four-stage pipeline that combined three learning sources: supervised fine-tuning on tool-calling trajectories, followed by directed distillation from three specialist teachers of 4 B parameters to restore reasoning abilities and instruction following that had regressed in the first stage, and then two rounds of reward-based reinforcement learning with transaction validation. This engineering reduced the rate of falling into repetitive reasoning loops from 22.9 % to 1.4 % and cut the number of tokens needed to answer the GPQA Diamond test by roughly threefold to 3,690 tokens, compared with about 11,563 tokens for the base model.

Knowing when an agent refrains from acting marks true maturity.BotPsy recorded 86.1 points on hallucination-detection tests in BFCL v4, the highest among all evaluated models, ensuring the agent declines execution when the appropriate tool is missing rather than fabricating random calls. In safety and malicious-attack tests, unsafe response rates fell to 21 % from roughly 36 % for the base model, and susceptibility to command-injection attacks dropped to 31 %.

The published paper clearly delineates the model’s limits: it is designed for low- to medium-complexity tasks such as calendar and email management, web browsing, and file control, and is unsuitable for high-risk critical tasks. General encyclopedic knowledge remains tied to parameter count, so the model is built to invoke external tools for factual retrieval rather than rely on internal memory, and its performance degrades with instruction accumulation and in low-resource languages compared with English. Tether released the model with open weights under the Apache 2.0 license for research and educational purposes, providing GGUF formats in three quantizations for use with llama.cpp and the QVAC software package.

A tangible shift in publishing calculations for technical teams in the Gulf, Egypt and the Levant.Data-localisation requirements in regional banking and government institutions impose strict limits on sending customer data to cloud model APIs abroad. An open model of this small size lets local development teams run automation agents for customer services and internal processes on private on-premise servers or even on employees’ portable devices without recurring cloud costs. However, the announced lag of the model in lower-resource languages means Arab developers cannot rely on it as a direct generative Arabic conversational model, but rather as an executive coordinator supplied with clear tool paths and precise documentation for data retrieval, shifting local skill demands from linguistic prompting to function-integration and API engineering.

Don't miss the next story

Subscribe for updates