What GPU do you need to run BotPsy-0.8B?

0.85B parameters (Hugging Face, accessed 2026-11-11). Q4_K_M GGUF: 0.53 GB (source). Architecture config (Hugging Face config.json): 6 attention layers (the rest are recurrent) × 2 KV heads × 256 head dim, so about 12 KB of FP16 KV cache per token of context (how we estimate context).

A 0.8B tool-calling (function-calling) assistant for on-device agents, from Tether AI Research (Tether Data, S.A. de C.V.), published under the qvac org. Fine-tuned from Qwen3.5-0.8B (huggingface.co/Qwen/Qwen3.5-0.8B), which has no page here. Tether says it suits tasks of low to medium difficulty, that it thinks before answering (enable_thinking=True; the chat template defaults to non-thinking), that English is recommended, and that it "is not intended for safety-critical or high-stakes automated decisions, or for applications involving money, personal information, health and other sensitive topics." It runs on Transformers, vLLM, llama.cpp and Tether's QVAC SDK. Text only, with no projector. The base model's MTP layers are not included (mtp_num_hidden_layers 0), so there is no MTP drafting. Q4_K_M is Tether's official imatrix-calibrated GGUF (botpsy-0.8b-q4_k_m-imat.gguf, 529,296,960 bytes); every size is in the table below. Release date is the Hugging Face repo creation date.