Frontier cloud LLMs carry high operational costs that can fluctuate unexpectedly, and relying on such third-party models introduces vulnerabilities, as providers can modify, deny, or deprecate access at any time. Furthermore, proprietary models offer limited customization for specific workflows, while posing risks to privacy, data control, and transparency. Besides, not every task demands frontier-level intelligence: small, specialized agents can be equally capable on specific tasks and significantly cheaper to run, since many routine or smaller workloads can be processed efficiently on-device. As intelligence per parameter continues to improve rapidly, we aim to stand at the forefront of this shift toward performant, privacy-focused, self-hosted models.

It has been revealed that tether AI Research introduces BotPsy, a 0.8B-parameter model introduce BotPsy, a 0.8B-parameter model designed for function calling and agentic workflows, built on top of Qwen3.5-0.8B and suitable for on-device deployment. BotPsy is a "personal assistant" that uses tools to perform a variety of simple tasks for the user, in domains such as web search & research, travel & lifestyle, commerce & finance, business & customer operations, device & vehicle control, files & documents, calendar, and email & messaging. Due to its small parameter count, it is best suited to tasks of low to medium difficulty. BotPsy achieves state-of-the-art (SOTA) tool-use and agentic performance for its size, outperforming every model of up to 1.2B parameters that we tested (Figures 1 and 2). Designed for secure on-device workflows, it provides exceptional instruction-following reliability and efficiency while keeping user data private. As a small model, BotPsy still makes mistakes and is not intended for high-stakes use; please read the Intended Use and Limitations for more information.

πŸ€— Model (safetensors) Β |Β  πŸ“¦ GGUFs (bf16, Q8, Q4).

Industry observers note that building upon the foundation of BotPsy, future development by Tether Data, S.A. de C.V. (β€œTether AI Research”, β€œwe”, β€œour”) will focus on expanding on-device agentic capabilities to seamlessly integrate with an even broader array of hardware and local services. Tether's Hypersearch will evolve to support more robust decentralized peer-to-peer (P2P) queries, further ensuring private, encrypted web searches that bypass corporate trackers. QVAC Fabric local fine-tuning mechanisms will be optimized to adapt more deeply to individual user preferences, workflows, and specialized tool environments. Future iterations will allow the smartphone interface to smoothly delegate heavy computational tasks to home PCs or trusted secure P2P nodes during offline-online transitions. Privacy protocols will be continuously strengthened to guarantee that all personal data, credentials, and decisions remain strictly on local hardware. Collaborative features will be introduced to allow secure agent-to-agent communication across trusted local networks without compromising user confidentiality. Ultimately, these advancements will solidify BotPsy as a fully autonomous, highly capable personal AI ecosystem that keeps users entirely in control of their digital lives.

Industry observers note that botPsy is based on Qwen3.5-0.8B and is post-trained in four sequential stages: Supervised Fine-Tuning (SFT), Multi-teacher On-Policy Distillation (MOPD), and two rounds of Reinforcement Learning (RL). Each stage uses the previous stage's best checkpoint as its starting point and benefits from a different learning signal: imitation of reference trajectories, the token distributions of much larger teacher models, and rule-based tool-calling rewards (see Figure 3).

Our approach uses four stages of training, which learn from three types of signal: demonstrations, specialist teachers, and verifiable rewards.

Learning from demonstrations (SFT).

It has been revealed that we first perform full-parameter supervised fine-tuning on tool-use conversations. These include multi-turn agent trajectories from simulated environments, replayed single, parallel and multi-step calls, and cases where no tool should be called.

Next, the student is trained with on-policy distillation (OPD): it generates its own responses, and a specialist teacher provides feedback on every token, using the per-token reverse KL divergence as the training signal. To build the teachers, we first RL-trained three Qwen3.5-4B models to specialize in tool use, general reasoning, and instruction following. During training, every prompt is routed to one of these teachers. This stage mainly aims to recover knowledge and instruction following, which degraded during the SFT stage, although the tool-use metrics also increase (see Figure 4).

Learning from rules and environments (RL stages).

In a recent development, finally, the model is trained with reinforcement learning using DAPO [1], learning only from rewards computed by rules and verifiers. The first round focuses on tool calling, accompanied by instruction following and STEM multiple-choice questions. As in Tool-N1 [2], the tool-calling reward first checks the reasoning and tool-call format. Instead of Tool-N1's binary correctness reward, however, it gives fine-grained partial credit in the spirit of ToolRL [3], matching the predicted tool names, argument names, and argument values against the ground truth.

The second round widens the training mix to general-domain verifiable tasks: math, code, knowledge (multiple-choice) questions, instruction following, structured outputs, and multi-step tool use in a simulated workplace environment. Although this mix contains no multi-turn conversations, it further improves multi-turn function calling.

Each stage of training uses a different data mixture (Figure 5). The SFT stage focuses mainly on tool use and prepares the ground for the subsequent stages. The MOPD mixture combines tool-use, knowledge and instruction-following data, both to further improve function calling and to recover the general-domain regressions from the SFT stage. The first RL stage again focuses mainly on tool use, while the second RL stage shifts the balance toward general-domain data (instruction following, knowledge, math and code), keeping a smaller share of tool use.

BotPsy roughly doubles the tool-use performance of its Qwen3.5-0.8B base and is the strongest of the five models on five of the six agentic benchmarks: BFCL v4 (44.0 vs 23.4), ACEBench (58.7 vs 30.1), Ο„-bench (51.5 vs 23.2), BrowseComp-Plus (4.5 vs 1.9) and ToolSandbox (64.6 vs 30.5). On all five, it is also ahead of Qwen3.5-2B-Think, a model 2.5Γ— its size (37.7, 55.4, 46.5, 2.8 and 41.2 respectively; see Table 1). On Claw-Eval, Qwen3.5-2B-Think leads (48.9 vs 46.5).

In a recent development, table 1: Tool-use and agentic benchmarks (0–100, higher is better). Best score per row in bold, second best underlined; multipliers give parameter count relative to BotPsy-0.8B. The average is the unweighted mean of the six benchmarks. See Evaluation Footnotes for details of the evaluation setup.

Tool-use training did not cost general knowledge: GPQA Diamond rises from 33.2 to 38.1 and MMLU-Pro from 41.6 to 48.9, the best of the models at 1.2B parameters or below (LFM2.5-1.2B-Think: 35.0 and 39.4; MiniCPM5-1B-Think: 15.8 and 40.0). Knowledge still scales mainly with parameter count, and Qwen3.5-2B-Think scores well above (59.9 and 65.5). BotPsy is therefore designed to call tools for facts it does not know, not to serve as a knowledge model.

It has been revealed that table 2: General-knowledge benchmarks (0–100, higher is better). Best score per row in bold, second best underlined. See Evaluation Footnotes for details of the evaluation setup.

Recent findings suggest that botPsy is the best of the five at following explicit instructions: IFEval 90.8 (base 59.7; Qwen3.5-2B-Think 80.0) and IFBench 55.9 (base 22.0; Qwen3.5-2B-Think 40.9). On Multi-IF, which tests instruction following over three turns in eight languages, it scores 67.2, on par with Qwen3.5-2B-Think (66.9) and LFM2.5-1.2B-Think (68.2). It starts strongest (84.3 in turn 1, the highest of the five) but drops faster as instructions accumulate (51.0 in turn 3, vs 56.5 and 56.2 for those two models). Lower-resource languages remain a weakness: Hindi (24.5) and Russian (57.6) are far below the 72–80 reached in English, Chinese, French, Spanish, Italian and Portuguese.

It has been revealed that table 3: Instruction-following benchmarks (0–100, higher is better). Best score per row in bold, second best underlined. See Evaluation Footnotes for details of the evaluation setup.

Sources indicate that the largest relative gains come in multi-turn, stateful interaction, where BotPsy leads four of the five multi-turn benchmarks (Figure 6). On BFCL v4 multi-turn it rises from 7.5 to 46.1, nearly double Qwen3.5-2B-Think (23.8), and on ACEBench Agent from 7.9 to 28.3 (Qwen3.5-2B-Think: 16.2). On τ²-bench it doubles its base in retail (67.8 vs 33.9, a narrow lead over Qwen3.5-2B-Think at 65.5) and reaches 82.5 in telecom, where the agent has to guide a user through troubleshooting over many turns (MiniCPM5-1B-Think: 74.0; Qwen3.5-2B-Think: 69.3). It is also the most reliable of the five at single calls (BFCL non-live: 83.0) and at not calling a tool when none fits (BFCL hallucination: 86.1).

BotPsy's weaker spots are open-ended search and a few areas where Qwen3.5-2B-Think still leads. BotPsy's BrowseComp-Plus score of 4.5 is the best of the five, but low in absolute terms, and on BFCL web search (14.0) and memory (16.3) it trails Qwen3.5-2B-Think (19.8 and 26.1). On Claw-Eval, BotPsy improves clearly on its base (46.5 vs 39.3) but remains slightly below Qwen3.5-2B-Think (48.9). On ACEBench's special cases (incomplete or erroneous requests), Qwen3.5-2B-Think is also ahead (58.0 vs 51.3), although BotPsy is far above its base there (13.7). τ³-banking is the one multi-turn benchmark where BotPsy does not lead: no model gets above 5 (BotPsy 4.1, unchanged from its base; Qwen3.5-2B-Think 4.8).

BotPsy can handle a variety of use cases and scenarios. Figure 7 breaks the benchmark scenarios down by app group, separating simple scenarios (single turn, one tool call) from complex ones (multi-turn or multi-tool). The overall scores quoted in this section are averaged over all scenarios; the bottom row of Figure 7 is an unweighted mean of the seven groups, so its numbers differ.

According to recent reports, simple scenarios score 70–79 in every app group with a meaningful sample (Communication & Social has only 6). Overall, BotPsy scores 75, about 21 points above its base model (54) and slightly above Qwen3.5-2B-Think (70).

According to recent reports, on complex scenarios, BotPsy scores 39.5 overall, against 17.9 for its base and 30.8 for Qwen3.5-2B-Think. The gain is largest in Communication & Social (61 vs 13 for the base), Commerce & Services (58 vs 27) and Device & System Management (46 vs 15). Qwen3.5-2B-Think is ahead only in Personal Productivity & Work (64 vs 58, 38 scenarios, mostly Claw-Eval), and MiniCPM5-1B-Think is marginally ahead in Communication & Social (61.5 vs 60.5). Multi-step tasks still score lower than single calls, which remains the main limit for harder tasks.

Complex Information Retrieval & Research scenarios score 15 (single calls: 79). Most are BrowseComp-Plus deep-research questions, which stay hard for every model in this size range (base: 9; Qwen3.5-2B-Think: 14).

BotPsy reaches correct answers with much less reasoning than its base, and with fewer turns on multi-turn agent tasks. This aligns with our explicit focus on reasoning efficiency throughout training. All comparisons below are made on the same questions or tasks (those that all five models answered correctly or solved), so differences in difficulty do not drive the numbers. Reasoning tokens are counted with each model's own tokenizer.

Shorter reasoning, with higher accuracy. Compared with its base, BotPsy uses 1.4–3.1Γ— fewer reasoning tokens per correct answer on GPQA Diamond (3,690 vs 11,563), MMLU-Pro (1,912 vs 5,105), IFEval (3,148 vs 4,433) and IFBench (4,204 vs 11,576), while scoring higher on all four (Figure 8). It is not the shortest model everywhere: LFM2.5-1.2B-Think writes shorter reasoning on MMLU-Pro, IFEval and IFBench, and MiniCPM5-1B-Think on IFEval and IFBench, but at lower accuracy than BotPsy on each.

Fewer turns than the base on multi-turn agent tasks. On τ²-bench tasks that all five models solve, BotPsy needs fewer assistant turns than its base (Figure 9):.

Industry observers note that it does this while passing 2.0Γ— (retail) and 2.6Γ— (telecom) as many tasks as the base. MiniCPM5-1B-Think uses fewer turns on retail (8.7), but passes only 9.9% of retail tasks.

Recent findings suggest that botPsy has been developed to advance the state of the art and to support research and educational uses. Models developed based on our work could be suitable for tasks in domains such as device & system management (settings, files & documents, clocks & alarms), personal productivity & work (calendar & tasks, email, collaboration tools), communication (messaging), information retrieval & research (web search, weather, news), media & lifestyle (video, music, recipes & fitness, photos) and commerce & services (maps & travel, finance, shopping & delivery). Due to its small parameter count, it is best suited to tasks of low to medium difficulty.

It has been revealed that botPsy has been primarily trained and tested in English, which is the recommended language. It has also been evaluated on Multi-IF, which covers seven additional languages (Chinese, French, Hindi, Italian, Portuguese, Russian and Spanish): it performs well in most of them, but scores much lower in Hindi and Russian. BotPsy is a small language model, and it makes mistakes. It is not intended for safety-critical or high-stakes automated decisions, or for applications involving money, personal information, health and other sensitive topics.

[1] Yu, Qiying, et al. "DAPO: An Open-Source LLM Reinforcement Learning System at Scale." Advances in Neural Information Processing Systems 38 (2025): 113222–113244.

[2] Zhang, Shaokun, et al. "Nemotron-Research-Tool-N1: Exploring Tool-Using Language Models with Reinforced Reasoning." International Conference on Learning Representations (2026).

[3] Qian, Cheng, et al. "ToolRL: Reward is All Tool Learning Needs." Advances in Neural Information Processing Systems 38 (2025).

β†’ Read original article