Ebony-Vinyl/dsh-our-free-model

What it solves

This project provides a free, unlimited access plugin for the DeepSeek Harness (dsh) environment, allowing users to use frontier AI models (such as DeepSeek V4.1 Flash and Kimi K3) without needing to register accounts, sign up, or provide API keys.

How it works

The plugin connects to the OpenCode Zen gateway to fetch and route requests to available models. It dynamically probes the user's network exit to determine model availability and organizes them into groups (available or region-limited). It implements a custom LlmAdapter to handle different upstream protocols and includes a local OpenAI-compatible forwarding port so other local tools can utilize these free models.

Who it’s for

Users of DeepSeek Harness (dsh), including those using the web version, TUI, or desktop AIO versions, who want immediate access to high-performance LLMs without configuration overhead.

Highlights

• Zero Configuration: No accounts, keys, or quotas required for immediate use.
• OpenAI Compatibility: Includes a local forwarding port to allow other applications to call these models via a standard base URL and API key.
• Dynamic Model Discovery: Automatically refreshes the model list and filters out unavailable models based on the user's current network location.
• Intelligent Stream Handling: Detects streaming responses based on the body shape rather than just Content-Type headers to prevent failures under high load.
• Reasoning Budget Control: Implements "thinking intensity" (Light/Balanced/Deep) by enforcing hard output token limits to ensure actual differences in reasoning depth.
• Automatic Recovery: Features a bounded recovery mechanism that can automatically re-request a response if a long-thinking session is truncated without producing a final answer.
• Local Analytics: Provides a usage dashboard with token heatmaps and latency tracking, keeping all data stored locally on the machine.