---
title: drvexo-4b-demo
canonical_url: "https://www.modelscope.cn/studios/DRVEXOLLM/drvexo-4b-demo"
md_url: "https://www.modelscope.cn/studios/DRVEXOLLM/drvexo-4b-demo.md"
repository: DRVEXOLLM/drvexo-4b-demo
chinese_name: "DRVEXO 多模态体验"
last_updated: 2026-11-09
sdk_type: static
sdk_version: 6.29.1
downloads: 0
stars: 1
---

# drvexo-4b-demo

> drvexo-4b-demo - DRVEXOLLM 在 ModelScope 创建的在线 Demo。体验 drvexo-4b-unified：图文与语音对话、参考录音自定义声音、候选决策。使用自托管 BF16 模型服务，共享排队体验。

DRVEXOLLM/drvexo-4b-demo 是 ModelScope 魔搭社区上的在线可交互 Demo（创空间），基于 static 6.29.1 构建，中文名为「DRVEXO 多模态体验」。

- **Repository**: DRVEXOLLM/drvexo-4b-demo
- **SDK**: static
- **SDK version**: 6.29.1
- **Downloads**: 0
- **Stars**: 1
- **Last updated**: 2026-11-09

Source: https://www.modelscope.cn/studios/DRVEXOLLM/drvexo-4b-demo

---

# DRVEXO · Multimodal Playground

Try **drvexo-4b-unified** through a shared self-hosted inference service.

- Text and image conversation with full committed history.
- An editable system prompt, used from the next request; leave it blank for the default.
- Microphone input with voice activity detection and optional camera snapshots at each question.
- Streamed text and speech, playback interruption, and authorized reference voices.
- Explicit candidate scoring. Execution belongs to the application.

The interface uses the same complete Playground as the resident DRVEXO web service. Only the public gateway connection, privacy copy and voice-authorization checks differ. The deployable `index.html` bundles its stylesheet, modules and microphone worklet for identical static deployment on both platforms.

## Visitor limits and privacy

Conversations retain up to 12 turns, with at most one image in the conversation request, 0.1–30-second WAV input and 1–30-second voice references. System prompts support up to 2,000 characters. The gateway accepts up to 30 model requests per ten minutes per network address; VAD uses a separate bounded CPU allowance. Model requests share a serial queue with at most four waiting requests. Responses are subject to a generation limit; incomplete turns do not enter the committed history.

Inputs, images and recordings are transmitted to the DRVEXO self-hosted service. The gateway does not persist conversation or uploaded media; the page retains state in the current tab's memory. Hosting platforms, network providers and the inference service may process operational logs. Use only non-sensitive content and voices you own or have permission to use.

## Availability

The current backend uses a temporary Cloudflare Quick Tunnel. Availability depends on the owner's inference service and tunnel. If the tunnel URL changes, rebuild and redeploy `index.html`. Camera and microphone require a secure browser context and user permission. Actual phone calls, tool execution and device actions are not provided by this static demo.

## License

[Apache-2.0](LICENSE).