A 0.8B agent that beats models 2.5x its size: Tether's BotPsy ships open weights and thinks 3x less

Tether AI Research's BotPsy-0.8B scores 45.0 across six agentic benchmarks against 38.8 for a model 2.5x its size, needs about 3x fewer reasoning tokens per correct answer, and ships Apache 2.0 weights plus GGUF builds - though every number is self-reported by the authors on their own harness.

A sub-billion-parameter agent that actually calls tools

On 2026-11-11, Tether AI Research (QVAC) published a model card and a blog post for BotPsy-0.8B: a 0.8B-parameter model built on Qwen3.5-0.8B and post-trained for function calling and agentic workflows. The stated target is on-device deployment, across domains that range from search and research to travel, commerce and finance, business operations, device and vehicle control, files, calendar, email and messaging.

One framing note before any numbers, because it governs how everything below should be read: every benchmark figure in this post is the authors’ own. They were measured on the authors’ own harness, on a single evaluation snapshot dated 2026-11-03, with no independent replication. These are vendor claims with unusually complete footnotes - not a verified leaderboard.

The headline: a 0.8B model ahead of a 2B model

Across six tool-use and agentic benchmarks, scored 0-100, BotPsy-0.8B reports:

Benchmark BotPsy-0.8B Base (Qwen3.5-0.8B)
BFCL v4 44.0 23.4
ACEBench 58.7 30.1
tau-bench 51.5 23.2
ToolSandbox 64.6 30.5
BrowseComp-Plus 4.5 1.9
Claw-Eval 46.5 39.3

The post’s own unweighted six-benchmark average is 45.0 for BotPsy-0.8B, 24.7 for the base, and 38.8 for Qwen3.5-2B-Think - a model 2.5x its size. BotPsy also leads the <=1.2B class the authors tested (Qwen3.5-0.8B-Think, LFM2.5-1.2B-Think, MiniCPM5-1B-Think) on all six benchmarks, and beats Qwen3.5-2B-Think on five of the six. The exception is Claw-Eval, where the 2B model scores 48.9 against BotPsy’s 46.5.

The recipe is the actual story

A smaller model winning on agentic tasks is not news by itself; what is transferable here is the four-stage training sequence the authors describe, and the fact that it uses three distinct feedback signals.

  1. Full-parameter SFT on tool-use conversations.
  2. Multi-Teacher On-Policy Distillation (MOPD), against three Qwen3.5-4B teachers, each reinforcement-learned for one specialty - tool use, general reasoning, instruction following - and each giving per-token reverse-KL feedback on the student’s own samples rather than on fixed prompts.
  3. Two rounds of DAPO reinforcement learning.

The detail worth stealing is the first RL round’s reward function. Rather than scoring a tool call binary, it checks reasoning and tool-call format first and then awards fine-grained partial credit for matching tool names, argument names and argument values separately, in the spirit of ToolRL. With sparse binary rewards, a small model gets almost no gradient signal for being nearly right; partial credit turns “wrong argument” into a smaller mistake that is still a gradient.

Multi-turn behaviour is where the relative gains sit

The base model is near-useless at multi-turn agentic work, and most of the improvement is concentrated there:

  • BFCL v4 multi-turn: 7.5 -> 46.1 (Qwen3.5-2B-Think: 23.8)
  • ACEBench Agent: 7.9 -> 28.3 (2B-Think: 16.2)
  • tau-squared-bench retail: 67.8 (2B-Think: 65.5)
  • tau-squared-bench telecom: 82.5 (2B-Think: 69.3)

The model also learns when not to call a tool: 86.1 on BFCL v4’s hallucination test, the highest of the five models the authors tested, and 51.3 on ACEBench special cases, up from 13.7 for its base.

Instruction following moves sharply too: IFEval 59.7 -> 90.8, and IFBench 22.0 -> 55.9, the best IFBench score among the five. Multi-IF lands at 67.2, on par with Qwen3.5-2B-Think (66.9) and LFM2.5-1.2B-Think (68.2).

Efficiency: the number that decides whether on-device is viable

This is arguably the more consequential result. On questions every model in the comparison answered correctly, BotPsy needs 1.4x to 3.1x fewer reasoning tokens per correct answer:

  • GPQA Diamond: 3,690 tokens vs 11,563 for the base
  • MMLU-Pro: 1,912 vs 5,105
  • IFEval: 3,148 vs 4,433
  • IFBench: 4,204 vs 11,576

The failure modes that matter for a phone-local agent also collapse: the repetitive-reasoning-loop rate falls from 22.9 percent to 1.4 percent, and runs that hit the 32k-token ceiling drop from 28.3 percent to 0.2 percent. On multi-turn agent tasks it also uses fewer assistant turns - tau-squared-bench retail 10.7 turns vs 11.7 (9 percent fewer), telecom 13.1 vs 17.6 (26 percent fewer) - while passing 2.0x (retail) and 2.6x (telecom) as many tasks as the base.

A 0.8B model that loops less is worth more than a 0.8B model that scores higher.

Where it still fails

The honest third story is the model’s ceiling. On complex multi-step information-retrieval and research scenarios it scores 15 out of 100, even though single-call retrieval in the same group scores 79. BFCL web search (14.0) and memory (16.3) trail Qwen3.5-2B-Think (19.8 and 26.1). On tau-cubed-banking no model the authors tested scores above 5, BotPsy included, at 4.1.

Read that as a scope statement, not a bug: this is a single-call and short-multi-step tool agent, not a research agent.

Knowledge also lags the 2.5x sibling by a wide margin - MMLU-Pro 48.9 against 65.5, GPQA Diamond 38.1 against 59.9 - even though both moved up from the base (41.6 and 33.2). The authors say plainly that BotPsy is meant to call tools for facts it does not know rather than serve as a knowledge model.

Safety, caveated

On the authors’ own automated red-teaming across six harm categories, BotPsy returns unsafe responses 21 percent of the time, against 36 percent for its base and lower than LFM2.5-1.2B-Think (35 percent) or MiniCPM5-1B-Think (47 percent). Under prompt-injection attacks it is at 31 percent versus 46 percent for the base. These percentages come from the authors’ own red-teaming, with no published standard behind them - directionally encouraging, not a safety guarantee.

The release terms are narrower than the benchmark table: the weights are Apache 2.0 for research and educational use, and the model card states it is not intended for high-stakes decisions or for money, personal information, health and other sensitive topics. Read the card’s terms before shipping it.

Getting it

The model card and the GGUF build are public on the Hub as of the post: safetensors in bf16, Q8 and Q4, plus GGUF files for the QVAC SDK and llama.cpp. The evaluation caveats from the post’s own footnotes are worth repeating: every benchmark using an external API was run on 2026-11-03, all models were evaluated in thinking mode, and several of those benchmarks are sensitive to provider-side changes in that window.

Authors: Milan Gritta, Jihye Back, JinUk Cho, Ohjoon Kwon, Patrik Lambert, Mathias Buus, Amril Nurman (Tether AI Research / QVAC).

Sources