Not tested by us yet. BotPsy-0.8B is a weights-only release with no hosted API and no OpenRouter listing, so our test harness cannot call it. Everything below comes from the QVAC announcement on Hugging Face, and every number is the vendor's. We'll update this post when a hosted endpoint appears or we run the GGUF ourselves.
BotPsy-0.8B is a 0.8 billion parameter model from Tether AI Research, fine-tuned from Qwen3.5-0.8B for function calling and agent loops on devices. On Tether's tool-use suite it averages 45.0, against 38.8 for the 2.5 times larger Qwen3.5-2B-Think and 24.7 for its own base. It is for small, cheap tool routing on a phone or laptop, not for knowledge questions.
| BotPsy-0.8B | |
|---|---|
| Maker | Tether AI Research, published through QVAC |
| Announced | 2026-11-11 |
| Base model | Qwen3.5-0.8B |
| Size | 0.8B parameters |
| Licence | Apache 2.0, stated for research and educational use |
| Formats | bf16, Q8, Q4 (safetensors and GGUF) |
| Access | Hugging Face weights and GGUF files for the QVAC SDK and llama.cpp |
| Price | Free to download, you pay for your own compute |
How it was trained
The announcement describes four stages. Supervised fine-tuning on tool-use conversations comes first, including multi-turn trajectories and cases where no tool should be called. Then multi-teacher on-policy distillation from three Qwen3.5-4B teachers, each reinforcement-trained for tool use, general reasoning or instruction following. Two reinforcement learning rounds follow. The first rewards tool names, argument names and argument values with partial credit. The second mixes math, code, knowledge, structured output and multi-step tool use.
The pattern is the point: a tiny base model gets most of its gains from teachers four times its size.
What the numbers say
All scores below are from Tether's Table 1, on a 0 to 100 scale, compared with its own base and three other small thinking models.
| Benchmark | BotPsy-0.8B | Qwen3.5-0.8B-Think (base) | Qwen3.5-2B-Think | LFM2.5-1.2B-Think | MiniCPM5-1B-Think |
|---|---|---|---|---|---|
| BFCL v4 | 44.0 | 23.4 | 37.7 | 28.1 | 22.8 |
| ACEBench | 58.7 | 30.1 | 55.4 | 39.4 | 24.5 |
| τ-bench | 51.5 | 23.2 | 46.5 | 6.5 | 28.8 |
| Claw-Eval | 46.5 | 39.3 | 48.9 | 31.6 | 31.0 |
| BrowseComp-Plus | 4.5 | 1.9 | 2.8 | 0.2 | 0.8 |
| ToolSandbox | 64.6 | 30.5 | 41.2 | 45.7 | 28.1 |
| Average | 45.0 | 24.7 | 38.8 | 25.3 | 22.7 |
BotPsy leads five of six columns and trails Qwen3.5-2B-Think on Claw-Eval. The gain over the base is largest on multi-turn work: BFCL v4 multi-turn goes from 7.5 to 46.1, against 23.8 for Qwen3.5-2B-Think. It also scores 90.8 on IFEval and 55.9 on IFBench, both above the 2B model.
The cost of the tuning shows up in knowledge. On GPQA Diamond BotPsy scores 38.1 against 59.9 for Qwen3.5-2B-Think, and on MMLU-Pro 48.9 against 65.5. Tether says as much: the model is meant to call tools for facts it does not know.
Limits the authors admit
- Difficulty: it suits tasks of "low to medium difficulty".
- Language: it was trained and tested mainly in English. Multi-IF drops to 24.5 in Hindi and 57.6 in Russian, from 72 to 80 in English, Chinese, French, Spanish, Italian and Portuguese.
- Search: BrowseComp-Plus is 4.5 and BFCL web search is 14.0 versus 19.8 for the 2B model.
- Banking agents: τ³-bench banking stays at 4.1, unchanged from the base.
- Sensitive use: Tether advises against high-stakes decisions or anything involving money, health or personal data.
The release also reports fewer reasoning loops (22.9% to 1.4% on IFBench) and 3,690 reasoning tokens on GPQA Diamond versus 11,563 for the base. Those are cost signals worth checking on your own prompts.
Who should care
- On-device agent builders. If your app needs a local model to pick a function and fill arguments, a 0.8B model with Q4 weights is a much smaller download than a 2B one, and the vendor's numbers say it is not worse at the job.
- Teams routing before a bigger model. A small local router that decides when to call a larger model fits the pattern in our guide to cheap-first routing with escalation, and BotPsy's hallucination score of 86.1 on declining when no tool fits is the number to verify first.
- Anyone comparing hosted options. If you would rather pay per call, see our ranking of the cheapest LLMs for function calling, and the explainer on how function calling works.
The fair reading today: a strong vendor claim for a model class that usually disappoints, backed by one lab's own suite. The test that matters is your own tool schemas, run locally, with the retry rate counted.