irfannkhan/GLM-5.3-Flash-demo

GLM-5.3-Flash — Multimodal Demo

[zai-org/GLM-5.3-Flash](https://huggingface.co/zai-org/GLM-5.3-Flash) — The first natively multimodal model in the GLM-5 series.

320B total / 18B active • hybrid sparse + linear attention • mHC architecture • 30T-token multimodal pre-training • 1/10th the price of GLM-5.2 • approaching Claude Opus 4.8 on coding/agentic benchmarks. Blog • Tech Report arXiv:2602.15763 • Z.ai API

This Space is a thin Inference Providers proxy — no model weights are hosted here. Inference is routed to hosted providers (Novita / Together / Baseten / Z.ai) via huggingface_hub.InferenceClient. Works on free accounts (ZeroGPU no-op).

Features

• —💬 Conversational chat with streaming, system prompt, temperature/topp/maxtokens
• —🖼️ Vision — upload an image + ask questions (image-text-to-text)
• —🧠 Reasoning control — reasoning_effort (low / high / max) + clear_thinking (GLM-5.3-Flash native)
• —🔐 Bring your own token — uses HF_TOKEN Space secret (creator-pays) if set, otherwise your personal HF token via Sign in (visitor-pays, inference-api scope)
• —⚡ Zero VRAM — runs on zero-a10g with a no-op GPU slot, so free users can host it (2-Space cap)

How to use

1. 1.(Optional) Sign in with Hugging Face if the Space has no HF_TOKEN secret — your token pays for inference (free tier available).
2. 2.Type a message, optionally attach an image, pick reasoning_effort and hit Send.
3. 3.Try the examples — vision, coding, and agentic prompts from the model card.

Reasoning docs

• —reasoning_effort: low / high / max (defaults to max if omitted). Pass explicitly for low/high.
• —clear_thinking: defaults to false in the template — set true for plain chat without exposed thinking trace.

Providers

Available via Inference Providers API: novita, together, baseten, zai-org (author). Default is novita. Override via Space variable INFERENCE_PROVIDER.

Check providers live:

curl "https://huggingface.co/api/models/zai-org/GLM-5.3-Flash?expand[]=inferenceProviderMapping"

Local run

pip install -r requirements.txt
HF_TOKEN=hf_... python app.py

Model license: MIT • 321B params • BF16/F8 • glm5_next arch • Conversational & image-text-to-text