AivexaNewsSearch
AI news for builders and product teamsChecked every hour

BotPsy: On-Device Personal Assistant with a Sub-Billion-Parameter Agentic Language Model

Collected Nov 11, 2026

Tether AI Research, part of Tether Data, S.A. de C.V., has introduced BotPsy, a 0.8B-parameter language model specialized for function calling and agentic workflows and intended for on-device deployment. It is built on Qwen3.5-0.8B and positioned as a "personal assistant" that drives tools across domains including 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 team frames the release as a bet that many routine tasks do not need frontier-level intelligence, and that a small, self-hosted model can be cheaper, more private and less exposed to a third-party provider changing or withdrawing access. BotPsy will ship with open weights under the Apache 2.0 license, stated for research and educational use, in three precisions (bf16, Q8, Q4), offered as Hugging Face safetensors and as GGUF files for the QVAC SDK and llama.cpp.

The headline numbers come from tool-use and agentic benchmarks. BotPsy roughly doubles its base model's performance on three central ones: 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, the last being a mean over the retail, telecom and banking subsets. The authors describe this as the best agentic performance in the 1.2B-parameter-or-smaller weight class, ahead of Qwen3.5-0.8B-Think, LFM2.5-1.2B-Think and MiniCPM5-1B-Think on all six benchmarks they used, the remaining three being ToolSandbox, Claw-Eval and BrowseComp-Plus.

More notable is the comparison against Qwen3.5-2B-Think, a model 2.5 times larger. BotPsy wins on five of the six: BFCL v4 at 44.0 versus 37.7, ACEBench at 58.7 versus 55.4, τ-bench at 51.5 versus 46.5, ToolSandbox at 64.6 versus 41.2, and BrowseComp-Plus at 4.5 versus 2.8. The larger model leads only on Claw-Eval, 48.9 against 46.5. This suggests that targeted post-training on tool use can buy more agentic capability at this scale than raw parameter count, at least on benchmark-style tasks.

Multi-turn, stateful interaction is where the relative gains are largest. On BFCL v4's multi-turn tasks BotPsy climbs from 7.5 to 46.1, nearly double Qwen3.5-2B-Think's 23.8, and on ACEBench Agent from 7.9 to 28.3 against 16.2 for the larger model. On τ²-bench, where the agent must hold a long customer conversation, follow a strict policy and act on the customer's behalf, it scores 82.5 in telecom, ahead of Qwen3.5-2B-Think at 69.3 and MiniCPM5-1B-Think at 74.0, and 67.8 in retail against 65.5. Grouped by application, its multi-step scores lead all tested models in commerce and services (58.3 versus 47.4) and device and system management (46.1 versus 31.9).

Knowing when not to act is treated as a first-class capability. On BFCL v4's hallucination tests, which check that the model declines when no suitable tool exists, BotPsy scores 86.1, the highest of the models tested. On ACEBench's special cases, where a request is incomplete, carries wrong parameters or cannot be served by the available tools, it reaches 51.3, up from 13.7 for its base, though Qwen3.5-2B-Think is ahead there at 58.0. For anyone building an agent that reads untrusted tool output, the safety figures are also relevant: automated red-teaming across six harm categories puts BotPsy's unsafe-response rate at 21%, against 36% for its base and 35% and 47% for LFM2.5-1.2B-Think and MiniCPM5-1B-Think respectively. Under prompt-injection attacks the unsafe rate falls from 46% to 31%.

Specializing for tools did not cost general skills, according to the reported results. IFEval rises from 59.7 to 90.8 and IFBench from 22.0 to 55.9, the highest IFBench score among the models tested, while GPQA Diamond moves from 33.2 to 38.1 and MMLU-Pro from 41.6 to 48.9. Knowledge still tracks parameter count, however: Qwen3.5-2B-Think scores 59.9 and 65.5 on those two, far above BotPsy. The team's own framing is that BotPsy is designed to call tools for facts it does not know rather than to serve as a knowledge model, which is a sensible division of labor for a small on-device agent backed by search.

Efficiency is a stated training focus, and the comparisons are made only on questions every model answered correctly so that difficulty does not skew them. On GPQA Diamond, BotPsy needs about 3,690 reasoning tokens against 11,563 for its base, roughly three times fewer, and 1,912 versus 5,105 on MMLU-Pro. On IFBench attempts, the rate of getting stuck in repetitive reasoning loops falls from 22.9% to 1.4%, and runs that hit the 32k-token ceiling drop from 28.3% to 0.2%. Shorter, more decisive reasoning translates directly into lower latency and battery cost on a phone, which is the whole point of an on-device assistant.

The training recipe has four sequential stages, each starting from the previous stage's best checkpoint. First comes full-parameter supervised fine-tuning on tool-use conversations, including multi-turn trajectories from simulated environments and replayed single, parallel and multi-step calls, plus cases where no tool should be called. Next is multi-teacher on-policy distillation: the student generates its own responses and three specialist Qwen3.5-4B teachers, each RL-trained for tool use, general reasoning or instruction following, score every token, with per-token reverse KL divergence as the signal. That stage mainly recovers knowledge and instruction following, which degraded during SFT. Two reinforcement learning rounds follow using DAPO with rule-based and verifier rewards: the first centers on tool calling with fine-grained partial credit for matching tool names, argument names and argument values, in the spirit of ToolRL, and the second widens to math, code, multiple-choice knowledge, instruction following, structured outputs and simulated workplace tasks. In practice, that staged mix is a plausible template for anyone post-training a small model for agent work.

The weak spots are worth weighing. Open-ended search remains hard: BrowseComp-Plus at 4.5 is best in class but low in absolute terms, and on BFCL web search (14.0) and memory (16.3) BotPsy trails Qwen3.5-2B-Think (19.8 and 26.1). Complex information retrieval and research scenarios average 15, against 79 for single calls in the same group. On Multi-IF, instruction following degrades as turns accumulate, starting at 84.3 in turn one and falling to 51.0 by turn three, and lower-resource languages are a clear gap, with Hindi at 24.5 and Russian at 57.6 versus 72–80 for English, Chinese, French, Spanish, Italian and Portuguese. τ³-banking stays below 5 for every model. The team also states plainly that BotPsy still makes mistakes and is not meant for high-stakes use—an important caveat for a model positioned to act on calendars, email and device controls.

Why it matters: developers weighing local agents now have a 0.8B option that claims to beat models twice its size on most agentic benchmarks, ships in quantized GGUF formats for llama.cpp, and is billed as cheap and private to run. Teams building assistants on phones or edge hardware should treat it as a candidate for low- to medium-difficulty tool workflows, with cloud models reserved for hard retrieval and high-stakes decisions—an inference on my part, but one the published weak spots support. The Apache 2.0 release also makes it straightforward to fine-tune on a specific tool environment, which is where small models tend to earn their keep.

Read at Hugging Face Blog

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt