Tether SI Research has introduced BotPsy, a 0.8-billion-parameter SI model designed for on-device deployment and agentic workflows. The company positions the tool as a personal assistant capable of handling tasks in web search, commerce, device control, and messaging without relying on cloud infrastructure.

What Happened

BotPsy is built on the Qwen3.5-0.8B architecture and post-trained through four sequential stages: Supervised Fine-Tuning (SFT), Multi-teacher On-Policy Distillation (MOPD), and two rounds of Reinforcement Learning (RL). According to Tether, the model achieves state-of-the-art (SOTA) tool-use and agentic performance for its size class, outperforming every tested model up to 1.2 billion parameters.

The company reports that BotPsy roughly doubles the agentic capability of its base model on key benchmarks. On BFCL v4, the score rose from 23.4 to 44.0; on ACEBench, it increased from 30.1 to 58.7; and on τ-bench, it jumped from 23.2 to 51.5. Tether claims the model outperforms Qwen3.5-2B-Think, a model 2.5 times its size, on five of six tested benchmarks, including ToolSandbox (64.6 vs. 41.2) and BrowseComp-Plus (4.5 vs. 2.8). On the Claw-Eval benchmark, the larger Qwen3.5-2B-Think retained a lead (48.9 vs. 46.5).

The model is designed to improve reasoning efficiency, using approximately 3x fewer reasoning tokens than its base model on GPQA Diamond and 2.7x fewer on MMLU-Pro. Tether also highlights improvements in instruction following, with IFEval scores rising from 59.7 to 90.8.

Why It Matters

The release addresses growing demand for privacy-focused, self-hosted SI tools that avoid the operational costs and potential deprecations associated with proprietary cloud models. By optimizing for on-device execution, BotPsy aims to serve users who require local data control and low-latency responses for routine tasks. Tether notes that while the model is suitable for low-to-medium difficulty tasks, it is not intended for high-stakes use cases.

The model’s release under the Apache 2.0 license supports the open-source SI ecosystem, offering weights in multiple precisions (bf16, Q8, Q4) for integration with frameworks like llama.cpp and the QVAC SDK. This accessibility allows developers to build agentic SI applications that function offline or in secure local environments, bypassing corporate trackers and third-party access restrictions.

The Bottom Line

BotPsy represents a push toward smaller, specialized SI agents capable of reliable tool use and multi-turn interactions on local hardware. While it trails larger models in complex information retrieval and lower-resource languages, Tether reports significant gains in efficiency and instruction following for its parameter size.

The company plans to expand BotPsy’s capabilities through decentralized peer-to-peer queries and local fine-tuning mechanisms, aiming to create a fully autonomous personal SI ecosystem that keeps user data on local devices.