The Pulse
Tether introduces BotPsy, a 0.8B on-device agent model
Tether AI Research has introduced BotPsy, a small model designed to use tools on a user's device. The team reports gains on agent benchmarks but says the model makes mistakes and is not intended for high-stakes decisions.

Tether AI Research has introduced BotPsy, a 0.8-billion-parameter model designed to use tools and run on a user's device, in a post published on Hugging Face. The model builds on Qwen3.5-0.8B and is intended for personal-assistant tasks ranging from web searches and calendars to files and messaging.
The developers say the model is best suited to low- to medium-difficulty tasks and can make mistakes. They describe it as a research and educational project rather than a model intended for high-stakes automated decisions.
Weights and intended uses
The model repository holds BF16 weights in Hugging Face safetensors format under the Apache 2.0 license. A linked repository contains GGUF files in BF16, Q8_0 and Q4_K_M formats for llama.cpp and the QVAC SDK.
In addition to productivity and research tasks, the team lists device and vehicle control, travel, commerce and customer operations among the areas where an assistant could use tools. Tether says BotPsy was primarily trained and tested in English, its recommended language.
Benchmark results
The team reports that BotPsy led the five models it tested on five of six tool-use and agent benchmarks, including comparisons with Qwen3.5-2B-Think. On BFCL v4, BotPsy scored 44.0, compared with 23.4 for its base model and 37.7 for the larger Qwen model; on ACEBench, the respective scores were 58.7, 30.1 and 55.4.
On Claw-Eval, the larger Qwen model scored 48.9 against BotPsy's 46.5, according to the published results. The team also reports that BotPsy scored 4.5 on BrowseComp-Plus, the highest of the five models tested but low in absolute terms.
The researchers grouped scenarios by application and complexity. They report an overall score of 75 for simple scenarios involving a single tool call, compared with 39.5 for complex scenarios involving multiple turns or tools; those overall scores were averaged over all scenarios.
Training and limitations
BotPsy underwent four sequential post-training stages: supervised fine-tuning, multi-teacher on-policy distillation and two rounds of reinforcement learning. The first stage used tool-use conversations, including simulated multi-turn interactions and cases where no tool should be called.
During distillation, the model generated responses while specialist teacher models provided token-level feedback. The reinforcement-learning stages used rewards calculated by rules and verifiers, with the latter stage adding tasks in areas including math, code, structured outputs and simulated workplace tool use.
In automated testing across six harm categories, the team found that BotPsy produced unsafe responses in 21% of cases, compared with 36% for its base model. Under prompt-injection attacks, the reported unsafe-response rate was 31%, compared with 46% for the base model.
Tether says the model makes mistakes and is not intended for safety-critical or high-stakes automated decisions, or applications involving money, personal information, health and other sensitive topics. Although it was also evaluated in seven additional languages, the team reports substantially lower instruction-following scores in Hindi and Russian than in the other languages tested.
Sources
Related stories
- Google adds offline speech input and playback to AI Edge Gallery
- AWS introduces cross-framework evaluation system for AI agents
- Plurai raises $13 million and opens agent simulation platform
- rakumo brings AI agent to invoice and estimate approval workflows
- Automattic launches Spacefast for agent-built apps