BotPsy-0.8B: Tether AI's On-Device Agentic Model Built on Qwen3.5
TL;DR: Tether AI Research has introduced BotPsy-0.8B, a 0.8B-parameter agentic language model built on Qwen3.5-0.8B for on-device tool use. It roughly doubles its base model's agentic benchmark scores and outperforms every model up to 1.2B parameters that the lab tested, making specialized local agents a practical alternative to frontier cloud models for low- to medium-difficulty tasks.
What BotPsy-0.8B is
BotPsy-0.8B is a sub-billion-parameter language model designed for function calling and agentic workflows. Tether AI Research built it on top of Qwen3.5-0.8B and positions it as a "personal assistant" that uses tools for tasks in web search and research, travel and lifestyle, commerce and finance, business and customer operations, device and vehicle control, files and documents, calendar, and email and messaging.
The model is intended for on-device deployment. Tether AI Research argues that frontier cloud LLMs carry high and variable operational costs, can be changed or deprecated by providers, offer limited workflow customization, and create privacy and data-control risks. It also states that not every task requires frontier-level intelligence: small, specialized agents can be cheaper and equally capable on specific tasks.
Because of its parameter count, BotPsy-0.8B is best suited to low- to medium-difficulty tasks. Tether AI Research explicitly says the model still makes mistakes and is not intended for high-stakes use.
Benchmark results: state of the art at 1.2B parameters and below
BotPsy-0.8B achieves state-of-the-art tool-use and agentic performance for its size, according to Tether AI Research. It outperforms every model of up to 1.2B parameters that the lab tested on all six tool-use and agentic benchmarks: BFCL v4, ACEBench, τ-bench, ToolSandbox, Claw-Eval, and BrowseComp-Plus.
Against its Qwen3.5-0.8B base model, BotPsy-0.8B roughly doubles agentic capability on key benchmarks: BFCL v4 rises from 23.4 to 44.0, ACEBench from 30.1 to 58.7, and τ-bench from 23.2 to 51.5. Its unweighted six-benchmark average is 45.0, compared with 24.7 for its base model, 38.8 for Qwen3.5-2B-Think, 25.3 for LFM2.5-1.2B-Think, and 22.7 for MiniCPM5-1B-Think.
| Benchmark | BotPsy-0.8B | Qwen3.5-0.8B-Think base | Qwen3.5-2B-Think | LFM2.5-1.2B-Think | MiniCPM5-1B-Think |
|---|---|---|---|---|---|
| BFCL v4 | 44.0 | 23.4 | 37.7 | 28.1 | 22.8 |
| ACEBench | 58.7 | 30.1 | 55.4 | 39.4 | 24.5 |
| τ-bench | 51.5 | 23.2 | 46.5 | 6.5 | 28.8 |
| Claw-Eval | 46.5 | 39.3 | 48.9 | 31.6 | 31.0 |
| BrowseComp-Plus | 4.5 | 1.9 | 2.8 | 0.2 | 0.8 |
| ToolSandbox | 64.6 | 30.5 | 41.2 | 45.7 | 28.1 |
| Average | 45.0 | 24.7 | 38.8 | 25.3 | 22.7 |
BotPsy-0.8B also beats Qwen3.5-2B-Think, a model 2.5× its size, on five of the six benchmarks: 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). Only on Claw-Eval does Qwen3.5-2B-Think lead, at 48.9 versus 46.5.
Multi-turn workflows and tool-call reliability
The largest relative gains are in multi-turn, stateful interaction. On BFCL v4 multi-turn tasks, BotPsy-0.8B rises from 7.5 for its base model to 46.1, nearly double Qwen3.5-2B-Think's 23.8. On ACEBench Agent, it rises from 7.9 to 28.3, compared with 16.2 for Qwen3.5-2B-Think.
On τ²-bench, which tests long customer conversations under strict policy, BotPsy-0.8B scores 67.8 in retail versus 33.9 for its base, and 82.5 in telecom versus 69.3 for Qwen3.5-2B-Think and 74.0 for MiniCPM5-1B-Think. It leads four of the five multi-turn benchmarks shown in the post. On τ³-bench banking, no model scores above 5; BotPsy-0.8B scores 4.1, unchanged from its base, while Qwen3.5-2B-Think scores 4.8.
BotPsy-0.8B is also designed to know when not to call a tool. On BFCL v4 hallucination tests, which check whether a model declines when no appropriate tool exists, it scores 86.1, the highest of all models tested. On ACEBench special cases involving incomplete requests, wrong parameters, or unavailable tools, it scores 51.3, up from 13.7 for its base model, though Qwen3.5-2B-Think is ahead at 58.0.
Knowledge and instruction following
Tool-use training did not cost general knowledge, according to the post. GPQA Diamond rises from 33.2 for the base model to 38.1 for BotPsy-0.8B, and MMLU-Pro rises from 41.6 to 48.9. Both are the best scores among models at 1.2B parameters or below that were tested. Knowledge still scales mainly with parameter count: Qwen3.5-2B-Think scores 59.9 on GPQA Diamond and 65.5 on MMLU-Pro. Tether AI Research therefore describes BotPsy-0.8B as a model designed to call tools for facts it does not know, not as a knowledge model.
Instruction following improves sharply. IFEval rises from 59.7 to 90.8, and IFBench rises from 22.0 to 55.9, the highest IFBench score of all models tested. On Multi-IF, which tests instruction following over three turns in eight languages, BotPsy-0.8B scores 67.2, on par with Qwen3.5-2B-Think at 66.9 and LFM2.5-1.2B-Think at 68.2. It starts strongest in turn 1 at 84.3 but drops faster as instructions accumulate, reaching 51.0 in turn 3 versus 56.5 and 56.2 for those two models.
Lower-resource languages remain a weakness. Hindi scores 24.5 and Russian 57.6, far below the 72–80 range reached in English, Chinese, French, Spanish, Italian, and Portuguese.
Safety and inference efficiency
In automated red-teaming across six harm categories, BotPsy-0.8B produces unsafe responses 21% of the time, compared with 36% for its base model. That is the lowest unsafe rate among models of up to 1.2B parameters tested: LFM2.5-1.2B-Think is at 35% and MiniCPM5-1B-Think at 47%. Under prompt-injection attacks, which matter for agents that read untrusted tool outputs, the unsafe rate drops from 46% for the base model to 31% for BotPsy-0.8B.
BotPsy-0.8B also reaches correct answers with less reasoning. On questions all five models answer correctly, it uses about 3× fewer reasoning tokens than its base model on GPQA Diamond (3,690 vs 11,563) and 2.7× fewer on MMLU-Pro (1,912 vs 5,105). Across all IFBench attempts, its rate of getting stuck in repetitive reasoning loops falls from 22.9% to 1.4%, and runs that hit the 32k-token limit fall from 28.3% to 0.2%. Tether AI Research says shorter, more decisive reasoning should reduce latency on-device.
On multi-turn agent tasks, BotPsy-0.8B needs fewer assistant turns than its base on τ²-bench tasks that all five models solve. On τ²-retail it uses 10.7 turns versus 11.7 for the base, 9% fewer, and on τ²-telecom it uses 13.1 turns versus 17.6, 26% fewer. It does this while passing 2.0× as many retail tasks and 2.6× as many telecom tasks as the base model.
Training methodology: four stages and three feedback signals
BotPsy-0.8B is post-trained in four sequential stages: supervised fine-tuning, multi-teacher on-policy distillation, and two rounds of reinforcement learning. Each stage starts from the previous stage's best checkpoint and uses a different learning signal: demonstrations, specialist teacher token distributions, and rule-based tool-calling rewards.
The first stage is full-parameter supervised fine-tuning on tool-use conversations, including multi-turn agent trajectories from simulated environments, replayed single, parallel, and multi-step calls, and cases where no tool should be called. The second stage uses on-policy distillation: the student generates its own responses, and a specialist teacher provides feedback on every token using per-token reverse KL divergence. The teachers are three Qwen3.5-4B models that were RL-trained separately for tool use, general reasoning, and instruction following. Each prompt is routed to one teacher. This stage mainly recovers knowledge and instruction following that degraded during supervised fine-tuning, while tool-use metrics also increase.
The final two stages use reinforcement learning with DAPO, learning only from rewards computed by rules and verifiers. The first RL round focuses on tool calling, along with instruction following and STEM multiple-choice questions. Its tool-calling reward first checks reasoning and tool-call format, then gives fine-grained partial credit by matching predicted tool names, argument names, and argument values against ground truth, in the spirit of ToolRL rather than Tool-N1's binary correctness reward. The second RL round widens the mix to general-domain verifiable tasks: math, code, knowledge questions, instruction following, structured outputs, and multi-step tool use in a simulated workplace environment. Although that mix contains no multi-turn conversations, it further improves multi-turn function calling.
Deployment, licensing, and limitations
The post states that BotPsy-0.8B will be released with open weights under the Apache 2.0 license for research and educational use. It will be available in three precisions—bf16, Q8, and Q4—as Hugging Face safetensors and as GGUF files for the QVAC SDK and llama.cpp.
BotPsy-0.8B has been primarily trained and tested in English, which is the recommended language. It is best suited to low- to medium-difficulty tasks and is not intended for safety-critical or high-stakes automated decisions, or for applications involving money, personal information, health, and other sensitive topics.
Its weak spots include open-ended search and a few areas where Qwen3.5-2B-Think still leads. BrowseComp-Plus scores 4.5, the best of the five models but low in absolute terms. On BFCL web search it scores 14.0 and on memory 16.3, trailing Qwen3.5-2B-Think at 19.8 and 26.1. On Claw-Eval it improves on its base, 46.5 versus 39.3, but remains below Qwen3.5-2B-Think at 48.9. Multi-step information retrieval is the clearest weak spot: complex Information Retrieval & Research scenarios score 15, compared with 79 for single calls in that group.
Broader vision
Tether AI Research says future development will expand on-device agentic capabilities across more hardware and local services. Planned work includes more robust decentralized peer-to-peer queries for Tether's Hypersearch, optimized local fine-tuning through QVAC Fabric, delegation of heavy computation from smartphones to home PCs or trusted secure P2P nodes during offline-online transitions, stronger privacy protocols, and secure agent-to-agent communication across trusted local networks.
What the release means
BotPsy-0.8B shows that a sub-billion-parameter model can be specialized for tool use and agentic workflows while preserving general knowledge and instruction following. Its benchmark profile—roughly double its base model's agentic scores, leadership among models up to 1.2B parameters, and competitive results against a 2.5× larger sibling on five of six agentic benchmarks—supports the argument that many routine assistant tasks can run locally rather than through frontier cloud APIs. The trade-offs are equally clear: weak open-ended search, lower-resource language gaps, and a stated ceiling on task difficulty.