Glonce

Tether AI Research introduces BotPsy, a 0.8B on-device agent model

modelsThe Glonce editorsHugging Face🔥61Glonce rating61Impact50Novelty60Relevance98Surprise25Players35Credibility80
Tether AI Research introduces BotPsy, a 0.8B on-device agent model

Tether AI Research has introduced BotPsy, a 0.8B-parameter model for function calling and agentic workflows, built on Qwen3.5-0.8B and meant to run on-device. The authors pitch it as a "personal assistant" that uses tools for simple tasks: 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. They say that because of its size it suits tasks of low to medium difficulty. The motivation they give is that frontier cloud models have operating costs that can swing, can be changed or withdrawn by providers, and pose privacy and control risks, while many routine tasks do not need frontier-level intelligence.

The headline claim is state-of-the-art tool-use and agentic performance for its size: BotPsy outperforms every model of up to 1.2B parameters that the authors tested (Qwen3.5-0.8B-Think, LFM2.5-1.2B-Think and MiniCPM5-1B-Think) on all six benchmarks: BFCL v4, ACEBench, τ-bench, ToolSandbox, Claw-Eval and BrowseComp-Plus. Against its base model, 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 (mean of the retail, telecom and banking subsets). ToolSandbox goes from 30.5 to 64.6 and BrowseComp-Plus from 1.9 to 4.5. BotPsy also tops Qwen3.5-2B-Think, a model 2.5× its size, on five of the six: 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 the 2B model lead (48.9 vs 46.5).

On multi-turn work, BotPsy scores 82.5 on the τ²-bench telecom domain, against 69.3 for Qwen3.5-2B-Think. On BFCL v4 multi-turn tasks it climbs from 7.5 to 46.1, nearly double Qwen3.5-2B-Think (23.8). Grouped by application, it leads in commerce and services (58.3 vs 47.4 for the next best, Qwen3.5-2B-Think) and in device and system management (46.1 vs 31.9). It also does better at declining when it should: on BFCL v4 hallucination tests, where no suitable tool exists, it scores 86.1, the highest of all models tested, and on ACEBench special cases (incomplete requests, wrong parameters, requests the tools cannot serve) it scores 51.3, up from 13.7 for the base model.

General skills improved too. IFEval rises from 59.7 to 90.8 and IFBench from 22.0 to 55.9, the highest IFBench score of all models tested. GPQA Diamond goes from 33.2 to 38.1 and MMLU-Pro from 41.6 to 48.9, best among models of 1.2B or fewer parameters (LFM2.5-1.2B-Think scores 35.0 and 39.4). Knowledge still scales mainly with parameter count, and Qwen3.5-2B-Think scores well above (59.9 and 65.5), so BotPsy is designed to call tools for facts it does not know rather than act as a knowledge model.

On safety, automated red-teaming across six harm categories gives BotPsy an unsafe-response rate of 21%, against 36% for its base model, the lowest among models up to 1.2B (LFM2.5-1.2B-Think 35%, MiniCPM5-1B-Think 47%). Under prompt-injection attacks the unsafe rate falls from 46% to 31%. On efficiency, on questions all five models answer correctly, BotPsy 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, repetitive reasoning loops fall from 22.9% to 1.4%, and runs hitting the 32k-token limit fall from 28.3% to 0.2%. The authors say shorter reasoning should reduce latency on-device.

Training has four sequential stages, each starting from the previous stage's best checkpoint. First comes full-parameter supervised fine-tuning on tool-use conversations: multi-turn agent trajectories from simulated environments, replayed single, parallel and multi-step calls, and cases where no tool should be called. Second is multi-teacher on-policy distillation (MOPD): the student generates its own responses and a specialist teacher gives per-token feedback using reverse KL divergence. The teachers are three Qwen3.5-4B models RL-trained for tool use, general reasoning and instruction following. This stage mainly recovers knowledge and instruction following that degraded during fine-tuning. The last two stages are reinforcement learning with DAPO, using only rule-based and verifiable rewards. Round one targets tool calling plus instruction following and STEM multiple-choice questions, with a tool-calling reward that checks format and then gives partial credit for matching tool names, argument names and argument values. Round two widens to math, code, knowledge questions, instruction following, structured outputs and multi-step tool use in a simulated workplace. That mix has no multi-turn conversations, yet it further improves multi-turn function calling.

Release: BotPsy will be released with open weights under the Apache 2.0 license for research and educational use, in three precisions (bf16, Q8, Q4), as Hugging Face safetensors and as GGUF files for the QVAC SDK and llama.cpp. A Broader Vision section lists plans by Tether Data, S.A. de C.V. (referred to as Tether AI Research): more robust decentralized peer-to-peer queries in Tether's Hypersearch, deeper local fine-tuning through QVAC Fabric, delegation of heavy tasks from phones to home PCs or trusted P2P nodes, and secure agent-to-agent communication across trusted local networks. No timescale is given.

Key facts

  • BotPsy is a 0.8B-parameter function-calling and agentic model post-trained from Qwen3.5-0.8B, aimed at on-device use for low to medium difficulty tasks.
  • The authors report it beats every tested model of up to 1.2B parameters on all six agentic benchmarks; BFCL v4 goes from 23.4 to 44.0 and τ-bench from 23.2 to 51.5 versus the base model.
  • It also beats Qwen3.5-2B-Think, a model 2.5× its size, on five of six benchmarks; only on Claw-Eval does the 2B model lead (48.9 vs 46.5).
  • Unsafe responses in red-teaming fall from 36% to 21%, and about 3× fewer reasoning tokens are used on GPQA Diamond (3,690 vs 11,563).
  • Training is SFT, multi-teacher on-policy distillation from three Qwen3.5-4B teachers, and two RL rounds; Apache 2.0 weights in bf16, Q8 and Q4 are stated as forthcoming.

Why it matters

Small models that can call tools are the route to assistants that run on a phone or laptop without a cloud provider. Tether AI Research reports that a 0.8B model can roughly double its base model's agentic scores and, on five of six benchmarks, beat a model 2.5× its size. Its stated case is cost, privacy and control: cloud pricing can fluctuate, access can be changed or withdrawn, and many routine workloads do not need frontier-level intelligence. The result is also a recipe: tool-use fine-tuning, distillation from larger specialist teachers, then reinforcement learning with verifiable rewards.

Who it affects

Developers building on-device or self-hosted assistants and agents, especially those who want a small model that works with llama.cpp or the QVAC SDK. Researchers working on small-model post-training will find a four-stage pipeline described in detail. Users of privacy-focused local assistants are the intended end audience, in domains such as calendar, email and messaging, files, commerce and device control.

How to use it

BotPsy is intended as a local tool-calling assistant for simple tasks. The authors say it will ship with open weights under Apache 2.0 for research and educational use, as Hugging Face safetensors and as GGUF files for the QVAC SDK and llama.cpp, in bf16, Q8 and Q4. Use it for low to medium difficulty tasks and have it call tools for facts rather than rely on its own knowledge: Qwen3.5-2B-Think scores well above it on GPQA Diamond and MMLU-Pro (59.9 and 65.5, against BotPsy's 38.1 and 48.9).

How solid is it

The post is a first-party announcement by Tether AI Research, and the benchmark results are the authors' own evaluations; no independent replication is mentioned. The comparisons are against a limited set of named models (Qwen3.5-0.8B-Think, Qwen3.5-2B-Think, LFM2.5-1.2B-Think, MiniCPM5-1B-Think), and the claim of state-of-the-art applies to models of up to 1.2B parameters that the authors tested. The reasoning-token comparison covers only questions all five models answer correctly. Reasoning-length and latency are linked only by the authors' expectation that shorter reasoning should reduce on-device latency.

Risks and caveats

The authors say BotPsy still makes mistakes and is not intended for high-stakes use, and that it suits only low to medium difficulty tasks. Absolute scores remain modest on some tests: BrowseComp-Plus is 4.5, and the best BFCL v4 score is 44.0. Safety is improved but not solved: 21% unsafe responses in red-teaming and 31% under prompt injection. Knowledge is limited by size. The release is described in future tense and limited to research and educational use, and the future features (Hypersearch P2P queries, delegation to home PCs, agent-to-agent communication) are plans with no timescale. The source gives no on-device latency, memory or hardware figures.

“As a small model, BotPsy still makes mistakes and is not intended for high-stakes use”

— Tether AI Research, BotPsy announcement