Upload README.md with huggingface_hub

README.md

@@ -54,14 +54,16 @@ You ⇄ Gradio chat ⇄ open LLM (HF Inference Providers, billed to your account
   pay-as-you-go applies *only if you have enabled it* (PRO accounts include monthly credits).
   Long agentic builds can make dozens of calls per turn; plan mode adds planner + specialist
   calls on top.
-- **Local model (`zerogpu-local/Qwen3.5-35B-A3B`):** runs
-  [Qwen3.5-35B-A3B](https://huggingface.co/Qwen/Qwen3.5-35B-A3B) (MoE, ~3B active params,
-  thinking disabled) on this Space's own ZeroGPU. No provider billing at all — but its
-  72 GB of bf16 weights need a **full-GPU `xlarge` slice**, which consumes your free ZeroGPU
-  quota at **2×** the normal rate (bigger quota when signed in, 8× with PRO). Each agent step
-  also pays a GPU attach + weight-upload cost, so turns are slower than the provider models,
-  and it remains weaker than the frontier options on complex scenes. Duplicators can override
-  with the `LOCAL_LLM_MODEL` / `LOCAL_LLM_SIZE` env vars (e.g. a small model on `large`).
+- **Local model (`zerogpu-local/Qwen3.5-35B-A3B-FP8`):** runs
+  [Qwen3.5-35B-A3B-FP8](https://huggingface.co/Qwen/Qwen3.5-35B-A3B-FP8) (MoE, ~3B active
+  params, official FP8 quantization) on this Space's own ZeroGPU. No provider billing at
+  all — its ~37 GB of FP8 weights fit the default half-GPU `large` slice (1× quota; bigger
+  quota when signed in, 8× with PRO). Each agent step still pays a GPU attach +
+  weight-upload cost, so turns are slower than the provider models, and it remains weaker
+  than the frontier options on complex scenes. A **💭 Thinking** toggle (local model only,
+  on by default) lets the model reason before each step — better results for extra GPU
+  seconds, never provider credits. Duplicators can override with the `LOCAL_LLM_MODEL` /
+  `LOCAL_LLM_SIZE` env vars (e.g. the bf16 checkpoint on `xlarge`).
 - **Renders / animations:** GPU renders always use ZeroGPU quota, never provider credits.
 - **For anyone duplicating this Space:** if you set an `HF_TOKEN` secret, visitors who are
   not signed in fall back to *that token* for inference — i.e. **your** account pays. Remove
@@ -93,6 +95,12 @@ not to integrate them for now:
 If image/video diffusion models (e.g. texture generation) are added to this Space later,
 AOKit is the right tool to reach for first.
 
+Also evaluated: [unsloth's Qwen3.5-35B-A3B checkpoints](https://huggingface.co/unsloth/Qwen3.5-35B-A3B-GGUF).
+Their quantized variants are GGUF (llama.cpp), which would replace the whole
+transformers + ZeroGPU-packing inference stack; their safetensors repo mirrors the
+official bf16 weights. The official FP8 checkpoint gives the size win inside the stack
+we already have, so that is what ships.
+
 ## Credits / licensing
 
 - `vendor/blender_mcp_addon/` and the `blender-mcp` Python package are the official