Supertron3-0.8B-GGUF

Supertron3-0.8B is a compact vision-language model from Suprem Org, fine-tuned from Qwen3.5-0.8B, purpose-built for GUI agents and agentic tool calling at the edge — interpreting visual interfaces across web, desktop, and CLI environments to emit precise pyautogui-style computer-use actions or valid JSON function calls. Its hybrid architecture interleaves Gated DeltaNet and Attention layers (24 layers, 1024 hidden dimension) with a vision encoder, retains a 262K native context window, and fits in a 1.7GB footprint suited for low-latency, on-device deployment. Despite being the smallest model in its evaluation set, Supertron3-0.8B ranks first on BFCL-style function calling (82% vs. 56% for its own Qwen3.5-0.8B base and 69% for the larger Qwen3.5-4B) and is the only model in the comparison that can reliably execute computer-use tasks at all — scoring 100% on Computer Use versus 0% for the unmodified base model — though it trails slightly on Mind2Web step accuracy (77% vs. 80% for the base), reflecting that the fine-tune specifically taught the base model to act rather than just converse. It's deployable via Transformers, vLLM, or SGLang, with known limitations around long-horizon multi-turn workflows and ScreenSpot-Pro-class grounding precision, which the authors attribute to the constraints of an 0.8B-parameter vision encoder; it's released under the Apache 2.0 license.

Model Files

llama.cpp

LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp