Apppublic

xariam/GLM-5.3-Flash-Uncensored

sourceHugging Faceupdated 26d agoView on Hugging Face

3likes

App README

GLM-5.3-Flash Uncensored — chat demo

Streaming chat (text + image input, optional reasoning trace) for `orcarouter/GLM-5.3-Flash-Uncensored-FP8`, the abliterated build of Z.ai's GLM-5.3-Flash (320B-total / 18B-active MoE, native vision, 1M context, block-FP8).

Why this Space is a client, not a host

The checkpoint is 306 GiB across 62 safetensors shards. No Space hardware tier can load it (ZeroGPU tops out at 96 GB; the largest dedicated tier would need multi-node serving), so the app talks to an OpenAI-compatible endpoint that serves the model instead:

—OrcaRouter — https://api.orcarouter.ai/v1 with an sk-orca-… key (the model's publisher). Check their model catalog for the exact model ID they expose.
—Your own server — vLLM or SGLang following the GLM-5.3-Flash serving recipes; the checkpoint is a byte-for-byte drop-in for the base model. With vLLM the default served name is the repo ID, orcarouter/GLM-5.3-Flash-Uncensored-FP8.

Configuration

Where	Name	Purpose
Space secret	`API_KEY`	Owner's key, used for every visitor when set
Space variable	`API_BASE_URL`	Default endpoint (falls back to OrcaRouter)
Space variable	`MODEL_ID`	Default model name (falls back to the HF repo ID)
Sidebar (per visitor)	API key / Base URL / Model ID	Bring-your-own endpoint; kept in the browser session only

The app runs on the zero-a10g flavor with a never-called @spaces.GPU no-op — no GPU time is ever requested, it just satisfies the platform's requirement for a decorated function.

Note from the model card

The publisher states this build "has had its safety alignment substantially removed", is released for interpretability / red-teaming / robustness research, and should not be deployed to end users without your own moderation and abuse-prevention layers. Leaving API_KEY unset (bring-your-own-key) is the simplest such layer for a public Space.