BotPsy: On-Gadget Private Assistant with a Sub-Billion-Parameter Agentic Language Mannequin

Frontier cloud LLMs carry excessive operational prices that may fluctuate unexpectedly, and counting on such third-party fashions introduces vulnerabilities, as suppliers can modify, deny, or deprecate entry at any time. Moreover, proprietary fashions supply restricted customization for particular workflows, whereas posing dangers to privateness, information management, and transparency. Apart from, not each job calls for frontier-level intelligence: small, specialised brokers might be equally succesful on particular duties and considerably cheaper to run, since many routine or smaller workloads might be processed effectively on-device. As intelligence per parameter continues to enhance quickly, we purpose to face on the forefront of this shift towards performant, privacy-focused, self-hosted fashions.

Tether AI Analysis introduces BotPsy, a 0.8B-parameter mannequin introduce BotPsy, a 0.8B-parameter mannequin designed for operate calling and agentic workflows, constructed on prime of Qwen3.5-0.8B and appropriate for on-device deployment. BotPsy is a “private assistant” that makes use of instruments to carry out quite a lot of easy duties for the consumer, in domains equivalent to internet search & analysis, journey & life-style, commerce & finance, enterprise & buyer operations, system & automobile management, information & paperwork, calendar, and e-mail & messaging. Because of its small parameter rely, it’s best suited to duties of low to medium problem. BotPsy achieves state-of-the-art (SOTA) tool-use and agentic efficiency for its dimension, outperforming each mannequin of as much as 1.2B parameters that we examined (Figures 1 and a pair of). Designed for safe on-device workflows, it gives distinctive instruction-following reliability and effectivity whereas holding consumer information non-public. As a small mannequin, BotPsy nonetheless makes errors and isn’t meant for high-stakes use; please learn the Meant Use and Limitations for extra data.

🤗 Mannequin (safetensors)  |  📦 GGUFs (bf16, Q8, This fall)

Copyright Complaints

We’ll take applicable actions in response to note of copyright infringement. For those who imagine your work has been used or copied in a way that infringes upon your mental property rights, please e-mail data-apps@tether.io figuring out and describing each the copyrighted work and alleged infringing content material to file a discover of infringement.

Key Highlights

Benchmark outcomes

• Twice the efficiency of its base mannequin: BotPsy roughly doubles the agentic functionality of Qwen3.5-0.8B-Suppose 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 (imply of the retail, telecom and banking subsets).
• Finest agent in its class: BotPsy outperforms main fashions within the ≤1.2B-parameter weight class (Qwen3.5-0.8B-Suppose, LFM2.5-1.2B-Suppose, MiniCPM5-1B-Suppose) on all six tool-use and agentic benchmarks: BFCL v4, ACEBench, τ-bench, ToolSandbox, Claw-Eval and BrowseComp-Plus.
• Outperforms its bigger sibling: BotPsy has the very best rating of all fashions examined on 5 of the six benchmarks, beating Qwen3.5-2B-Suppose, a mannequin 2.5× its dimension, on 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). Solely on Claw-Eval does Qwen3.5-2B-Suppose lead (48.9 vs 46.5).

Mannequin capabilities

• Extra dependable in multi-turn workflows: On τ²-bench, the place the agent should maintain an extended buyer dialog, observe a strict coverage, and act on the shopper’s behalf, it scores 82.5 in telecom, forward of Qwen3.5-2B-Suppose (69.3). On BFCL v4’s multi-turn duties, it rises from 7.5 to 46.1 over its base, almost double Qwen3.5-2B-Suppose (23.8). Grouped by software, its multi-step scores lead all fashions examined in commerce and providers (58.3 vs 47.4 for the subsequent greatest, Qwen3.5-2B-Suppose) and in system and system administration (46.1 vs 31.9).
• Is aware of when (not) to name instruments: An agent that acts should additionally know when to not act. On BFCL v4’s hallucination assessments, which verify that the mannequin declines when no applicable instrument exists, BotPsy scores 86.1, the very best of all fashions examined. On ACEBench’s particular circumstances, the place the request is incomplete, accommodates unsuitable parameters, or can’t be served by the given instruments, it scores 51.3, up from 13.7 for its base mannequin.
• Follows directions whereas preserving information: BotPsy makes a speciality of instrument use whereas preserving basic expertise. Instruction following improves sharply, with IFEval rising from 59.7 to 90.8 and IFBench from 22.0 to 55.9, the very best IFBench rating of all fashions examined. Common information additionally improves, with GPQA Diamond rising from 33.2 to 38.1 and MMLU-Professional from 41.6 to 48.9. On all 4 benchmarks, BotPsy outperforms each different mannequin of 1.2B parameters or fewer that we examined.

Security, effectivity and deployment

• Safer than main fashions in its class: In automated red-teaming throughout six hurt classes, BotPsy produces unsafe responses 21% of the time, in contrast with 36% for its base mannequin. That is the bottom fee of all fashions of as much as 1.2B parameters (LFM2.5-1.2B-Suppose: 35%; MiniCPM5-1B-Suppose: 47%). Below prompt-injection assaults, which matter most for brokers that learn untrusted instrument outputs, the unsafe fee drops from 46% for its base mannequin to 31%.
• Extra environment friendly reasoning and fewer looping: On questions that each one 5 fashions reply accurately, BotPsy reaches the reply with about 3× fewer reasoning tokens than its base mannequin on GPQA Diamond (3,690 vs 11,563) and a pair of.7× fewer on MMLU-Professional (1,912 vs 5,105). Throughout all IFBench makes an attempt, its fee of getting caught in repetitive reasoning loops falls from 22.9% to 1.4%, and runs that hit the 32k-token restrict fall from 28.3% to 0.2%. Shorter, extra decisive reasoning ought to cut back latency on-device.
• • Open-source and able to use: BotPsy shall be launched with open weights underneath the Apache 2.0 license for analysis and academic use. It is going to be out there in three precisions (bf16, Q8, This fall), as Hugging Face safetensors and as GGUF information for the QVAC SDK and llama.cpp.

Broader Imaginative and prescient

Constructing upon the inspiration of BotPsy, future growth by Tether Information, S.A. de C.V. (“Tether AI Analysis”, “we”, “our”) will concentrate on increasing on-device agentic capabilities to seamlessly combine with a good broader array of {hardware} and native providers. Tether’s Hypersearch will evolve to help extra strong decentralized peer-to-peer (P2P) queries, additional guaranteeing non-public, encrypted internet searches that bypass company trackers. QVAC Material native fine-tuning mechanisms shall be optimized to adapt extra deeply to particular person consumer preferences, workflows, and specialised instrument environments. Future iterations will enable the smartphone interface to easily delegate heavy computational duties to residence PCs or trusted safe P2P nodes throughout offline-online transitions. Privateness protocols shall be constantly strengthened to ensure that each one private information, credentials, and choices stay strictly on native {hardware}. Collaborative options shall be launched to permit safe agent-to-agent communication throughout trusted native networks with out compromising consumer confidentiality. In the end, these developments will solidify BotPsy as a completely autonomous, extremely succesful private AI ecosystem that retains customers fully in charge of their digital lives.

Methodology

BotPsy relies on Qwen3.5-0.8B and is post-trained in 4 sequential phases: Supervised Wonderful-Tuning (SFT), Multi-teacher On-Coverage Distillation (MOPD), and two rounds of Reinforcement Studying (RL). Every stage makes use of the earlier stage’s greatest checkpoint as its place to begin and advantages from a unique studying sign: imitation of reference trajectories, the token distributions of a lot bigger instructor fashions, and rule-based tool-calling rewards (see Determine 3).

Coaching process: 4 phases, three varieties of suggestions

Our method makes use of 4 phases of coaching, which be taught from three varieties of sign: demonstrations, specialist lecturers, and verifiable rewards.

Studying from demonstrations (SFT)

We first carry out full-parameter supervised fine-tuning on tool-use conversations. These embody multi-turn agent trajectories from simulated environments, replayed single, parallel and multi-step calls, and circumstances the place no instrument needs to be known as.

Studying from lecturers (MOPD)

Subsequent, the scholar is skilled with on-policy distillation (OPD): it generates its personal responses, and a specialist instructor gives suggestions on each token, utilizing the per-token reverse KL divergence because the coaching sign. To construct the lecturers, we first RL-trained three Qwen3.5-4B fashions to concentrate on instrument use, basic reasoning, and instruction following. Throughout coaching, each immediate is routed to one among these lecturers. This stage primarily goals to recuperate information and instruction following, which degraded in the course of the SFT stage, though the tool-use metrics additionally enhance (see Determine 4).

Studying from guidelines and environments (RL phases)

Lastly, the mannequin is skilled with reinforcement studying utilizing DAPO [1], studying solely from rewards computed by guidelines and verifiers. The primary spherical focuses on instrument calling, accompanied by instruction following and STEM multiple-choice questions. As in Software-N1 [2], the tool-calling reward first checks the reasoning and tool-call format. As a substitute of Software-N1’s binary correctness reward, nonetheless, it offers fine-grained partial credit score within the spirit of ToolRL [3], matching the anticipated instrument names, argument names, and argument values in opposition to the bottom reality.

The second spherical widens the coaching combine to general-domain verifiable duties: math, code, information (multiple-choice) questions, instruction following, structured outputs, and multi-step instrument use in a simulated office surroundings. Though this combine accommodates no multi-turn conversations, it additional improves multi-turn operate calling.

Information composition

Every stage of coaching makes use of a unique information combination (Determine 5). The SFT stage focuses primarily on instrument use and prepares the bottom for the following phases. The MOPD combination combines tool-use, information and instruction-following information, each to additional enhance operate calling and to recuperate the general-domain regressions from the SFT stage. The primary RL stage once more focuses primarily on instrument use, whereas the second RL stage shifts the stability towards general-domain information (instruction following, information, math and code), holding a smaller share of instrument use.

Outcomes

Software use & brokers

BotPsy roughly doubles the tool-use efficiency of its Qwen3.5-0.8B base and is the strongest of the 5 fashions on 5 of the six agentic benchmarks: BFCL v4 (44.0 vs 23.4), ACEBench (58.7 vs 30.1), τ-bench (51.5 vs 23.2), BrowseComp-Plus (4.5 vs 1.9) and ToolSandbox (64.6 vs 30.5). On all 5, it’s also forward of Qwen3.5-2B-Suppose, a mannequin 2.5× its dimension (37.7, 55.4, 46.5, 2.8 and 41.2 respectively; see Desk 1). On Claw-Eval, Qwen3.5-2B-Suppose leads (48.9 vs 46.5).

Desk 1: Software-use and agentic benchmarks (0–100, larger is healthier). Finest rating per row in daring, second greatest underlined; multipliers give parameter rely relative to BotPsy-0.8B. The common is the unweighted imply of the six benchmarks. See Analysis Footnotes for particulars of the analysis setup.

Data

Software-use coaching didn’t price basic information: GPQA Diamond rises from 33.2 to 38.1 and MMLU-Professional from 41.6 to 48.9, the very best of the fashions at 1.2B parameters or under (LFM2.5-1.2B-Suppose: 35.0 and 39.4; MiniCPM5-1B-Suppose: 15.8 and 40.0). Data nonetheless scales primarily with parameter rely, and Qwen3.5-2B-Suppose scores properly above (59.9 and 65.5). BotPsy is due to this fact designed to name instruments for details it doesn’t know, to not function a information mannequin.

Desk 2: Common-knowledge benchmarks (0–100, larger is healthier). Finest rating per row in daring, second greatest underlined. See Analysis Footnotes for particulars of the analysis setup.

Instruction following

BotPsy is the very best of the 5 at following express directions: IFEval 90.8 (base 59.7; Qwen3.5-2B-Suppose 80.0) and IFBench 55.9 (base 22.0; Qwen3.5-2B-Suppose 40.9). On Multi-IF, which assessments instruction following over three turns in eight languages, it scores 67.2, on par with Qwen3.5-2B-Suppose (66.9) and LFM2.5-1.2B-Suppose (68.2). It begins strongest (84.3 in flip 1, the very best of the 5) however drops quicker as directions accumulate (51.0 in flip 3, vs 56.5 and 56.2 for these two fashions). Decrease-resource languages stay a weak point: Hindi (24.5) and Russian (57.6) are far under the 72–80 reached in English, Chinese language, French, Spanish, Italian and Portuguese.

Desk 3: Instruction-following benchmarks (0–100, larger is healthier). Finest rating per row in daring, second greatest underlined. See Analysis Footnotes for particulars of the analysis setup.

Multi-turn conversations

The biggest relative features are available multi-turn, stateful interplay, the place BotPsy leads 4 of the 5 multi-turn benchmarks (Determine 6). On BFCL v4 multi-turn it rises from 7.5 to 46.1, almost double Qwen3.5-2B-Suppose (23.8), and on ACEBench Agent from 7.9 to twenty-eight.3 (Qwen3.5-2B-Suppose: 16.2). On τ²-bench it doubles its base in retail (67.8 vs 33.9, a slender lead over Qwen3.5-2B-Suppose at 65.5) and reaches 82.5 in telecom, the place the agent has to information a consumer via troubleshooting over many turns (MiniCPM5-1B-Suppose: 74.0; Qwen3.5-2B-Suppose: 69.3). It is usually probably the most dependable of the 5 at single calls (BFCL non-live: 83.0) and at not calling a instrument when none suits (BFCL hallucination: 86.1).

Weak spots

BotPsy’s weaker spots are open-ended search and some areas the place Qwen3.5-2B-Suppose nonetheless leads. BotPsy’s BrowseComp-Plus rating of 4.5 is the very best of the 5, however low in absolute phrases, and on BFCL internet search (14.0) and reminiscence (16.3) it trails Qwen3.5-2B-Suppose (19.8 and 26.1). On Claw-Eval, BotPsy improves clearly on its base (46.5 vs 39.3) however stays barely under Qwen3.5-2B-Suppose (48.9). On ACEBench’s particular circumstances (incomplete or inaccurate requests), Qwen3.5-2B-Suppose can also be forward (58.0 vs 51.3), though BotPsy is much above its base there (13.7). τ³-banking is the one multi-turn benchmark the place BotPsy doesn’t lead: no mannequin will get above 5 (BotPsy 4.1, unchanged from its base; Qwen3.5-2B-Suppose 4.8).

The place it really works: app-level efficiency breakdown

BotPsy can deal with quite a lot of use circumstances and situations. Determine 7 breaks the benchmark situations down by app group, separating easy situations (single flip, one instrument name) from advanced ones (multi-turn or multi-tool). The general scores quoted on this part are averaged over all situations; the underside row of Determine 7 is an unweighted imply of the seven teams, so its numbers differ.

Single instrument calls work properly in every single place

Easy situations rating 70–79 in each app group with a significant pattern (Communication & Social has solely 6). Total, BotPsy scores 75, about 21 factors above its base mannequin (54) and barely above Qwen3.5-2B-Suppose (70).

The biggest features are in multi-step duties

On advanced situations, BotPsy scores 39.5 general, in opposition to 17.9 for its base and 30.8 for Qwen3.5-2B-Suppose. The acquire is largest in Communication & Social (61 vs 13 for the bottom), Commerce & Providers (58 vs 27) and Gadget & System Administration (46 vs 15). Qwen3.5-2B-Suppose is forward solely in Private Productiveness & Work (64 vs 58, 38 situations, largely Claw-Eval), and MiniCPM5-1B-Suppose is marginally forward in Communication & Social (61.5 vs 60.5). Multi-step duties nonetheless rating decrease than single calls, which stays the principle restrict for tougher duties.

Multi-step data retrieval is the weak spot

Complicated Info Retrieval & Analysis situations rating 15 (single calls: 79). Most are BrowseComp-Plus deep-research questions, which keep onerous for each mannequin on this dimension vary (base: 9; Qwen3.5-2B-Suppose: 14).

Inference effectivity

BotPsy reaches right solutions with a lot much less reasoning than its base, and with fewer activates multi-turn agent duties. This aligns with our express concentrate on reasoning effectivity all through coaching. All comparisons under are made on the identical questions or duties (people who all 5 fashions answered accurately or solved), so variations in problem don’t drive the numbers. Reasoning tokens are counted with every mannequin’s personal tokenizer.

Shorter reasoning, with larger accuracy. In contrast with its base, BotPsy makes use of 1.4–3.1× fewer reasoning tokens per right reply on GPQA Diamond (3,690 vs 11,563), MMLU-Professional (1,912 vs 5,105), IFEval (3,148 vs 4,433) and IFBench (4,204 vs 11,576), whereas scoring larger on all 4 (Determine 8). It isn’t the shortest mannequin in every single place: LFM2.5-1.2B-Suppose writes shorter reasoning on MMLU-Professional, IFEval and IFBench, and MiniCPM5-1B-Suppose on IFEval and IFBench, however at decrease accuracy than BotPsy on every.

Fewer turns than the bottom on multi-turn agent duties. On τ²-bench duties that each one 5 fashions clear up, BotPsy wants fewer assistant turns than its base (Determine 9):

• τ²-retail: 10.7 turns, in opposition to 11.7 for the bottom (9% fewer) and 10.6 for Qwen3.5-2B-Suppose.
• τ²-telecom: 13.1 turns, in opposition to 17.6 for the bottom (26% fewer) and 13.8 for Qwen3.5-2B-Suppose (5% fewer).

It does this whereas passing 2.0× (retail) and a pair of.6× (telecom) as many duties as the bottom. MiniCPM5-1B-Suppose makes use of fewer activates retail (8.7), however passes solely 9.9% of retail duties.

Meant Use and Limitations

BotPsy has been developed to advance the state-of-the-art and to help analysis and academic makes use of. Fashions developed based mostly on our work could possibly be appropriate for duties in domains equivalent to system & system administration (settings, information & paperwork, clocks & alarms), private productiveness & work (calendar & duties, e-mail, collaboration instruments), communication (messaging), data retrieval & analysis (internet search, climate, information), media & life-style (video, music, recipes & health, pictures) and commerce & providers (maps & journey, finance, procuring & supply). Because of its small parameter rely, it’s greatest suited to duties of low to medium problem.

BotPsy has been primarily skilled and examined in English, which is the beneficial language. It has additionally been evaluated on Multi-IF, which covers seven extra languages (Chinese language, French, Hindi, Italian, Portuguese, Russian and Spanish): it performs properly in most of them, however scores a lot decrease in Hindi and Russian. BotPsy is a small language mannequin, and it makes errors. It isn’t meant for safety-critical or high-stakes automated choices, or for purposes involving cash, private data, well being and different delicate subjects.

Quotation

@misc{botpsy2026,
  title  = {BotPsy: On-Gadget Private Assistant with a Sub-Billion-Parameter Agentic Language Mannequin},
  writer = {{Tether Information, S.A. de C.V. d.b.a. Tether AI Analysis}},
  yr   = {2026},
  be aware   = {Hugging Face weblog}
}

References

[1] Yu, Qiying, et al. “DAPO: An Open-Supply LLM Reinforcement Studying System at Scale.” Advances in Neural Info Processing Methods 38 (2025): 113222–113244.

[2] Zhang, Shaokun, et al. “Nemotron-Analysis-Software-N1: Exploring Software-Utilizing Language Fashions with Bolstered Reasoning.” Worldwide Convention on Studying Representations (2026).

[3] Qian, Cheng, et al. “ToolRL: Reward is All Software Studying Wants.” Advances in Neural Info Processing Methods 38 (2025).

Analysis Footnotes

• All scores are percentages (larger is healthier).
• All fashions have been evaluated in considering (reasoning) mode.
• All benchmarks that use an exterior API have been run on November 3, 2026.
• ¹ BFCL v4: general rating calculated utilizing the official weighted method (40% Agentic, 30% Multi-Flip, 10% Reside, 10% Non-Reside, and 10% Hallucination), averaged over 3 runs.
• ² ACEBench: general accuracy. All fashions have been evaluated in function-calling (FC) mode; imply of two runs.
• ³ τ-bench: common of τ²-bench (retail, telecom) and τ³-bench (banking), excluding airline; imply of three runs.
• ⁴ Claw-Eval: official common job rating, the place every job scores security × (0.8 × completion + 0.2 × robustness); imply of three runs.
• ⁵ BrowseComp-Plus: accuracy, with a most of 30 steps for all fashions.
• ⁶ ToolSandbox: imply similarity rating over 8 settings.
• ⁷ GPQA Diamond: accuracy over all 198 questions, imply of three runs, considering enabled.
• ⁸ MMLU-Professional: accuracy.
• ⁹ IFEval, IFBench: prompt-level strict accuracy.
• ¹⁰ Multi-IF: common of three runs.