README.md

Agent Zero — ZeroGPU Native

This repo contains the reference implementation for the Agent Zero ZeroGPU Space.

➡️ Live Space

The fully operational Space is at:
https://huggingface.co/spaces/ScottzillaSystems/agent-zero-orchestration

Architecture

• SDK: Gradio (required for ZeroGPU)
• Hardware: ZeroGPU (H200, 70GB VRAM)
• Decorator: @spaces.GPU(duration=180) for all inference functions
• Models: Loaded on-demand per request via transformers

Models (ScottzillaSystems Fleet)

Model	Tier	Size	Architecture	Repo
ChatGPT-5	T0	494M	Qwen2ForCausalLM	ScottzillaSystems/ChatGPT-5
Qwen3.5 9B Opus	T1	9.6B	Qwen3_5ForConditionalGeneration	ScottzillaSystems/Huihui-Qwen3.5-9B-Claude-4.6-Opus-abliterated
SuperGemma4	T1	7.5B	Gemma4ForConditionalGeneration	ScottzillaSystems/supergemma4-e4b-abliterated
Cydonia 24B	T2	24B	MistralForCausalLM	ScottzillaSystems/Cydonia-24B-v4.1
Qwen3.6 27B	T3	27.8B	CausalLM	ScottzillaSystems/Huihui-Qwen3.6-27B-abliterated
Qwen3 VL 8B	VL	8.8B	ConditionalGeneration	ScottzillaSystems/Huihui-Qwen3-VL-8B-Instruct-abliterated
Qwen3.5 9B Base	T1	9.6B	Qwen3_5ForConditionalGeneration	ScottzillaSystems/Qwen3.5-9B

Key Design Decisions

1. ZeroGPU requires Gradio SDK — Docker SDK is not supported
2. Models load inside @spaces.GPU — GPU is allocated per-request
3. AutoModelForImageTextToText for multimodal models (Qwen3.5, SuperGemma4, Qwen3 VL)
4. AutoModelForCausalLM for standard text models (ChatGPT-5, Cydonia, Qwen3.6 27B)
5. Smart auto-routing with fallback chain (T3→T2→T1→T0)
6. No LiteLLM, no TGI, no Docker Compose — pure transformers + ZeroGPU

Setup

1. Set HF_TOKEN as Space Secret
2. Set hardware to ZeroGPU in Space Settings
3. Done — models load on first request