huggingface-local-models: Run Hugging Face GGUF Models Locally with Skill
huggingface-local-models is an Agent Skill in the Hugging Face skills repository that guides local model selection and inference on CPU, Metal, CUDA, or ROCm using llama.cpp and GGUF. The default workflow includes searching the Hub with apps=llama.cpp, reviewing quantization recommendations on the local-app page, confirming filenames via the tree API, and launching llama-cli or llama-server with the -hf option, using --hf-repo/--hf-file when needed or converting from Transformers. The article covers installation and enabling, key considerations for quantization selection, a smoke test for an OpenAI-compatible local service, and usage notes.
Read Moredsh-wsl-llamacpp: A Plugin for Integrating with Local llama.cpp / Unsloth Desktop
`dsh-wsl-llamacpp` is a plugin for DeepSeek Harness (DSH) designed to simplify access to local LLM inference services. The plugin connects to local llama.cpp or Unsloth Desktop via an OpenAI-compatible interface, provides the `llama_status` and `llama_chat` tools, and supports reusing GGUF models. The installation command is `dsh plugin add github:173787247/dsh-wsl-llamacpp`. Users can configure the baseUrl, default model, and timeout. The plugin runs with the current process's permissions and is suitable for scenarios that require managing local WSL environment inference flows within DSH.
Read Moreiyicode/dsh-ctx-probe: Automatically Detect Local Model Context
The dsh-ctx-probe plugin resolves 400 overflow issues caused by configuration mismatches in the DSH llm-pi-ai local server. It probes the actual context windows of llama.cpp and Ollama servers in real time, enabling bidirectional automatic synchronization with local settings. To prevent probe storms, it uses single-flight and exponential backoff strategies, and supports real-time probing for local requests and caching for remote requests. After installation, restart DSH Web Host; use the `/ctx-probe force` command to view the configuration, effective settings, and probed values.
Read MoreRunning Local Models in DeepSeek Harness: vmarcelo49/dsh-local-models
DeepSeek Harness (DSH) introduces a new `dsh-local-models` plugin that enables direct invocation of local `.gguf` models within the web interface. Designed to work out of the box, the plugin provides model selection, context and KV cache tuning, real-time VRAM estimation, configuration persistence, and OpenAI-compatible routing. After installation, users must restart DSH to load models through the interface and register them as LLM providers, with support for automatic startup and idle eviction.
Read Moredsh-llamacpp: Local llama.cpp Adapter
`dsh-llamacpp` is an adapter plugin designed specifically for DeepSeek Harness (DSH) to connect to a local llama.cpp server. It communicates through an OpenAI-compatible interface, registers as the `llamacpp` provider in Harness, and supports streaming output and text-based tool calling. The plugin allows the server address, model name, and API key to be configured via environment variables or a configuration file. It currently supports text only, does not support images or Ollama, and includes robust error handling.
Read MoreIntroduction to the dsh-llama-model-manager Plugin
dsh-llama-model-manager is a DeepSeek Harness plugin designed specifically for the Windows environment, aimed at resolving the issues of cumbersome model switching, chaotic process management, and resource leaks during local deployment of large models. Through a fixed OpenAI-compatible gateway, it automatically manages the lifecycle of `llama-server`, including model loading, unloading, graceful shutdown, and VRAM release. It supports operations via a visual interface, provides process status monitoring and log viewing, and eliminates the need for frequent configuration changes, offering developers a stable and efficient local inference solution.
Read Moredsh-model-pin: Limit model request scope to prevent llama.cpp router reloads
The DeepSeek Harness (DSH) plugin dsh-model-pin resolves the “model roulette” problem, preventing local model reloads, cache loss, and routing errors caused by requests to disallowed models. The plugin enforces an allowlist at the resolution layer, forcing mismatched requests to be redirected or rejected, ensuring that sub-agents and background tasks always use the specified model. It is suitable for resource-constrained environments and scenarios requiring strict routing control, and can be easily enabled and customized via configuration files.
Read MoreIntroduction to DeepSeek Harness Local Model Plugin
`dsh-plugin-local-model` is a plugin for DeepSeek Harness designed to simplify local model inference. It uses a dedicated settings page to automatically launch llama.cpp upon the first conversation message after a model has been configured, and it automatically unloads after 5 minutes of inactivity to release resources. The plugin also supports one-click switching between parameter presets, ensuring offline operation and file security. Users must prepare their own models and tools; for installation, it is recommended to refer to the source code. This tool is suitable for developers who need local deployment, privacy protection, and frequent parameter switching.
Read MoreDeepSeek Harness Plugin: dsh-llama-responses
The DeepSeek Harness plugin dsh-llama-responses adopts a plugin-based architecture that routes sub-agent requests to a local llama.cpp service over the OpenAI protocol, reducing costs and protecting privacy. It supports multi-turn memory, streaming tool calls, and sub-agent isolation. During installation, make sure llama.cpp is running; once the adapter is configured, sub-agents can use a local model (e.g., qwen3.8-27b). Enabling the sub-agent tool in the configuration file allows the main agent and local sub-agents to operate independently.
Read Moredsh-local-llm-controller: Integrating Local llama.cpp Capabilities for DSH
The `dsh-local-llm-controller` plugin is designed to address the cumbersome process of deploying large models locally in DeepSeek Harness (DSH), enabling one-click start/stop and configuration management through its “everything is a plugin” approach. It supports A/B dual-slot management, switching between text and vision modes, and 8 parameter presets, and is automatically registered after installation. Users must prepare `llama-server` and model files themselves; after configuring the directories and paths in the plugin, they can start the service with one click. The plugin is currently compatible only with the DSH RC branch, does not include the server itself, and its vision mode requires specific build-version support.
Read More