Oleksandr Kuzmenko, EditorPublished 5 min readAI-assisted · editor-reviewed
TL;DR
- Scores 44.0 on BFCL v4 and 51.5 on τ-bench, outperforming models up to 1.2B and 2.5× larger checkpoints.
- Cuts reasoning loop failures from 22.9% to 1.4% to optimize battery and compute on mobile and desktop chips.
- Distributed under Apache 2.0 with bf16, Q8, and Q4 GGUF builds for llama.cpp.
Key facts
- Parameter count
- 0.8B
- License
- Apache 2.0
- Base model
- Qwen3.5-0.8B
- BFCL v4 score
- 44.0 (vs 23.4 base)
- Reasoning loop failure rate
- 1.4% (down from 22.9%)
Source: Key facts — Parameter count
Source: Key facts. Parameter count: 0.8B. Each figure is copied from the story.
- Parameter count0.8B
Performance and Benchmark Gains
Tether's BotPsy-0.8B doubles the tool-calling capability of its Qwen3.5-0.8B-Think foundation across several multi-turn and agent evaluations:
- BFCL v4: Climbs from 23.4 to 44.0 overall, and jumps from 7.5 to 46.1 on multi-turn subsets.
- ACEBench: Increases from 30.1 to 58.7 (reaching 51.3 on malformed or incomplete tool parameter edge cases).
- τ-bench: Averages 51.5 across retail, telecom, and banking (scoring 82.5 in telecom customer workflows).
- Safety and Guardrails: Reduces automated red-teaming unsafe responses from 36% down to 21%, with prompt-injection vulnerabilities declining from 46% to 31%.
Reasoning Efficiency and Quantization
Unlike verbose reasoning models that slow down edge inference, BotPsy achieves answers with fewer generated tokens: approximately 3× fewer reasoning tokens on GPQA Diamond (3,690 vs 11,563) and 2.7× fewer on MMLU-Pro compared to its base checkpoint. Context overflow runs hitting the 32k-token ceiling fell from 28.3% to 0.2%.
Weights are available under Apache 2.0 as Hugging Face safetensors and pre-quantized GGUF binaries (bf16, Q8, and Q4) compatible with llama.cpp and the QVAC SDK.
Try it in 2 minutes
llama-cli -hf qvac/botpsy -m botpsy-0.8b-q8_0.gguf -p "<|im_start|>system\nYou are a helpful assistant with tool calling abilities.<|im_end|>"When to use
- On-device personal assistants running on embedded hardware, consumer laptops, or mobile phones.
- Privacy-critical tool-dispatching pipelines requiring zero data transit to external cloud APIs.
When NOT to use
- High-stakes production tasks requiring advanced mathematics, large codebase refactoring, or zero error tolerance.
- Workflows with deeply accumulated, multi-page instruction prompts where 0.8B context retention degrades.
What to do today
- Download the BotPsy Q4 or Q8 GGUF files to test local function calling on low-spec edge instances.
- Implement tool-validation fallbacks utilizing BotPsy's high tool-refusal accuracy (86.1 on BFCL v4).
Step names are taken from What to do today.
Step names are taken from What to do today. Download the BotPsy Q4 or Q8 GGUF files to test local function calling on low-spec edge instances.. Implement tool-validation fallbacks utilizing BotPsy's high tool-refusal accuracy (86.1 on BFCL v4).
- Download the BotPsy Q4 or Q8 GGUF files to test local function calling on low-spec edge instances.
- Implement tool-validation fallbacks utilizing BotPsy's high tool-refusal accuracy (86.1 on BFCL v4).