MorningOpen sourceNotable

Tether releases BotPsy, a 0.8B open model for on-device tool calling

Tether AI Research says its Qwen3.5-based model roughly doubles its base on agent benchmarks and beats a 2B model on five of six, by its own testing.

Promptea Editorial3 min read

Tether AI Research, the AI unit of stablecoin issuer Tether, today published BotPsy-0.8B, a 0.8-billion-parameter model trained specifically for function calling and multi-step agent work on phones and laptops. The weights are on Hugging Face under an Apache 2.0 license, in BF16 safetensors plus GGUF builds at BF16, Q8_0 and Q4_K_M for llama.cpp. The release is accompanied by a long technical write-up with benchmarks and the training recipe.

The pitch is narrow and explicit: not every agent step needs a frontier model. A small model that reliably picks the right tool, fills in the right arguments and declines when no tool fits can run locally, cheaply and without sending user data to a cloud API. Tether says BotPsy is best suited to tasks of low to medium difficulty and is not meant to be a knowledge model.

What Tether reports

All numbers below come from Tether's own evaluation, run in thinking mode, with external-API benchmarks run on November 3. They have not been independently reproduced.

  • Roughly double its base model on tool use. Against Qwen3.5-0.8B-Think, BFCL v4 goes from 23.4 to 44.0, ACEBench from 30.1 to 58.7 and a combined τ-bench score from 23.2 to 51.5.
  • Ahead of a model 2.5× its size on five of six agent benchmarks. BotPsy beats Qwen3.5-2B-Think 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). Qwen3.5-2B-Think still leads on Claw-Eval (48.9 vs 46.5).
  • Better at knowing when not to call a tool. It scores 86.1 on BFCL v4's hallucination tests, the highest of the five models compared.
  • Much stronger instruction following. IFEval rises from 59.7 to 90.8 and IFBench from 22.0 to 55.9.
  • Shorter reasoning. On questions every model answered correctly, it used about 3,690 reasoning tokens on GPQA Diamond versus 11,563 for its base, and reasoning loops on IFBench fell from 22.9% to 1.4% of attempts.

How it was trained

BotPsy starts from Qwen3.5-0.8B and goes through four stages: supervised fine-tuning on tool-use conversations, on-policy distillation from three specialist Qwen3.5-4B teachers (tool use, reasoning, instruction following), and two rounds of reinforcement learning with DAPO on rule-checked rewards. The tool-call reward gives partial credit for matching tool names, argument names and argument values instead of a simple pass/fail. Tether says the distillation stage was needed mainly to recover the general knowledge and instruction following that fine-tuning had degraded, a useful detail for anyone trying to specialize small models.

The weak spots, by Tether's own account

The write-up is unusually candid about limits. Open-ended search remains poor: BrowseComp-Plus at 4.5 is the best of the group but low in absolute terms, and BotPsy trails Qwen3.5-2B-Think on BFCL's web-search and memory subsets. General knowledge still scales with size, so GPQA Diamond sits at 38.1 against 59.9 for the 2B model. Instruction following degrades as turns accumulate, and scores in Hindi (24.5) and Russian (57.6) are far below the 72-80 range in English, Spanish and other major languages. In Tether's automated red-teaming, unsafe responses dropped from 36% to 21%, and under prompt injection from 46% to 31%. That is an improvement, but it still means roughly one in three injected attacks succeeded.

There is also a licensing wrinkle worth noting. The repository files are under a standard Apache 2.0 license, which permits commercial use, but the blog frames the model as developed "to support research and educational uses" and warns against using it for money, personal data, health or other sensitive applications. Teams considering it for products should read the license and the intended-use section together.

Why it matters

For developers building agents, the practical question is routing: which steps can a cheap local model handle, and which need a large hosted one. BotPsy is a concrete data point that a sub-1B model, trained with the right mix of distillation and verifiable rewards, can handle single tool calls and structured multi-turn flows reasonably well, while still failing at open-ended research. The claims are vendor-reported, so the sensible next step is to test it against your own tool schemas before trusting it with anything that acts on a user's behalf.

Why this matters

  • Small local models that call tools reliably let developers route routine agent steps away from expensive hosted models.
  • The training recipe, which pairs distillation from specialist teachers with partial-credit tool-call rewards, is a reusable pattern for specializing small models.
  • All benchmark gains are self-reported, and the model's own documentation flags prompt-injection and high-stakes use as risks.

Key takeaways

  • BotPsy-0.8B is out now on Hugging Face under Apache 2.0, with GGUF builds for llama.cpp.
  • Tether reports BFCL v4 at 44.0 versus 23.4 for its Qwen3.5-0.8B base, and wins over Qwen3.5-2B-Think on five of six agent benchmarks.
  • Open-ended web search, general knowledge and non-English languages such as Hindi and Russian remain weak.
  • Test it on your own tool schemas before deploying; the blog positions it for research and education, not sensitive applications.

Sources

  1. Tether AI Research (Hugging Face blog)Primary
    BotPsy: On-Device Personal Assistant with a Sub-Billion-Parameter Agentic Language Model
    huggingface.co
  2. Tether AI Research (Hugging Face)Primary
    qvac/BotPsy-0.8B model card
    huggingface.co
  3. Tether AI Research (Hugging Face)Primary
    qvac/BotPsy-0.8B-GGUFs
    huggingface.co

How we choose and verify sources

Tether BotPsy-0.8B: open on-device tool-calling model · Promptea