DeepSeek Harness Plugin: dsh-wsl-vllm
`dsh-wsl-vllm` is a plugin for DeepSeek Harness (DSH) designed to simplify access to and interaction with vLLM services in WSL. It integrates vLLM's OpenAI-compatible interface, using port 8000 by default, into the DSH ecosystem and can be installed via the `dsh plugin` command. The plugin provides `vllm_status` to check service status and model loading state, and `vllm_chat` for chat completion. It supports Docker execution and configuration via environment variables. Ensure that your DSH version is ≥ 0.1.2 and that you have access to the service.
Read Moredsh-wsl-docker: WSL Docker Engine Diagnostics and vLLM Container Checks
The DeepSeek Harness (DSH) plugin `dsh-wsl-docker` is used to automatically diagnose the status of Docker and vLLM containers in WSL environments. It provides a `docker_doctor` tool that supports checking Docker paths, daemon configuration, context, and vLLM container operation. Developers can quickly troubleshoot local model deployment environments with simple commands, reducing debugging time. The plugin requires DSH version ≥ 0.1.2, can be installed via a command, and is available for use after restarting the service.
Read Moredsh-vision-3090-fix: A Local Proxy Plugin to Limit the Number of Images per Request
`dsh-vision-3090-fix` is a DeepSeek Harness plugin designed to resolve 400 errors caused by the vLLM backend’s default one-image-per-request limit conflicting with DSH’s internal byte-based adapter limit. It runs as a local reverse proxy, intercepts requests, replaces excess images with text placeholders, and enforces a configurable per-request image limit (default: 1), while preserving streaming response support. After installation, configure the proxy parameters and update the Provider’s `baseURL`. The tool is suitable for self-hosted vision model environments using vLLM with `--limit-mm-per-prompt image=1`, offering a low-cost, non-invasive solution.
Read MoreDeepSeek Harness Plugin: vLLM-Ascend Profiler Analyzer
The DeepSeek Harness plugin is designed to address the performance profiling challenges encountered when deploying vLLM on Ascend NPUs. By parsing profiling artifacts generated by CANN and torch_npu, it supports streaming analysis and visualization, automatically distinguishes between the Prefill and Decode stages, and generates optimization recommendations and reports. The plugin has zero runtime dependencies and can be installed via pnpm or a local link.
Read MoreIntroduction to the dsh-llm-openai-completions Plugin
`dsh-llm-openai-completions` is an adapter in DeepSeek Harness for connecting to custom gateways, primarily resolving compatibility issues related to role definitions, thinking mode, and visual input. Its core features include enforcing the use of the system role, separating thinking content, supporting image uploads, and adapting to various Qwen model formats. Users can install it via the DSH command line and enable it in their configuration file. Please note that maintenance for this project was discontinued in October 2026, remains at version 0.1.0, and the 0.2.0 version with visual features was never released.
Read MoreDeepSeek Harness Local Qwen Model Adapter (dsh-llm-qwen-local)
`dsh-llm-qwen-local` is a plugin-centric framework for deploying Qwen models locally and integrating them into the DSH ecosystem. The plugin provides an OpenAI-compatible interface through a vLLM backend, supports multimodal toggles, inference level configuration, and image projection, and features zero runtime dependencies along with a visual Web configuration interface. To use it, DSH version 0.1.2-rc.1 or later is required, and vLLM startup parameters must be configured correctly. The plugin is released under the MIT License and aims to simplify the process of integrating local models.
Read MoreMoonshot Releases Full Weights of Kimi K3: The First 3T-level Open Source Model
On July 27, 2026, Moonshot AI released the full weights of Kimi K3 on Hugging Face and GitHub: it has a total of 2.8T parameters, 104B activations, and a 1M context window, adopting MoE + KDA attention and MXFP4 native quantization. This article sorts out the architecture specifications, coding evaluation performance and vLLM self-hosting key points, and discusses the gap between open-source weights and closed-source cutting-edge models as well as the deployment threshold.
Read MoreQuickly Deploy a DeepSeek-R1 Service from Scratch
Here are the simplest commands to introduce how to deploy the DeepSeek-R1 service. Anaconda is assumed to be already installed, and the vllm framework is used, making it easy to deploy even in China.
Read More