First run
Download the appropriate launcher from the download page. On Windows, right-click the ZIP, extract it completely into a folder you control, and run Setup.cmd from inside that folder. Running files directly inside a compressed ZIP can fail.
The setup checks prerequisites. Follow the on-screen prompts, and keep the final console output if a step fails. The launcher source ZIP does not include all engines or model weights. A working Python installation, supported drivers and sufficient free storage may be required.
For Linux, use the dedicated Linux repository instructions. The exact shell commands and prerequisites are maintained with its code.
Choose an engine
| Engine | Use it for | Check first |
|---|---|---|
| Strata | Its supported MoE architectures and expert offloading workflow. | Exact supported checkpoint, prepared data, disk/RAM requirements and engine release. |
| llama-server | Compatible dense or MoE GGUF models, using the native llama.cpp server. | Architecture support, backend build, context and offload settings. |
| Ollama / LM Studio | Models those applications support through their local APIs. | The selected model is loaded and its local endpoint is reachable. |
Strata and llama-server are alternative execution paths. Installing one does not make every model compatible with it. Avoid running two large engines simultaneously until you have measured your available memory.
Choose a model
Use the model library to understand the checkpoint, task, license and runtime. The catalog identifies source-reviewed candidates; it does not guarantee successful installation or speed on your computer.
Start with a small model, a short prompt and a modest context. Increase one setting at a time. Large model downloads can require additional prepared data and temporary disk space beyond the advertised artifact size.
Connect your tools
DeepSeek Harness is a coding client and agent workspace. It can use a supported local model endpoint; its name does not mean that it must run a DeepSeek model. The model engine does inference, while the client controls the conversation and available tools.
YOUR CODING CLIENT
DeepSeek Harness / supported CLI client
│
├── local model API ── Strata or llama-server
│
└── approved MCP bridge ── Unreal / Blender / Android toolsMCP bridges expose capabilities from an application to an agent. Review each bridge's permissions and choose a specific project folder. A local model is not a sandbox: the executing tools determine what it can read, change or send over a network.
Presets & settings
The preset registry exports inspectable JSON. Its website schema is a configuration template and may need mapping to your installed launcher version. It does not execute a command or automatically import into every launcher.
Context and output budgets share the model's window. A useful preflight rule is:
allowed output ≤ server context − actual prompt tokens − safety bufferUse the backend's tokenizer where supported. A character-count approximation is not a guarantee against context errors. Sampling options and KV-cache formats must be supported by the selected engine.
Memory & context
Saved project memory is selected information that can be included in future requests. It does not expand the model's native context. The launcher's optional memory policy controls which stored notes it includes, with an off setting and user-verified modes.
A response that reaches its output budget may need a bounded continuation. Tool calls and structured output need special handling. Never assume a hidden continuation can safely complete every coding task.
Updates & recovery
Back up your settings and projects first. Review release notes and keep a working copy until the new version starts correctly. Plugins and custom profiles should live outside replaceable application files.
Download from the official release or repository and inspect any changed dependencies. The website does not silently update your computer or run downloaded presets.
Troubleshooting
The engine stays on “starting”
Check its console log, whether the process is still initializing, and whether its configured health endpoint responds. First loads and data preparation can take time. Confirm free RAM, VRAM and disk space before retrying.
Generation is slower than a published benchmark
Compare exact model, quantization, context, cache state, runtime version and competing applications. Prompt-processing speed differs from generation speed. See the lab's measured conditions.
Setup reports Python was not found
A Windows Store execution alias alone is not a working Python installation. Use the current setup's prerequisite checks or follow the repository's first-run guide. Save the exact error rather than repeatedly launching overlapping installations.
Community accounts
This site supports email/password registration and a separate public profile. Messages are public, and uploaded profile images are public. Administrators can review reports through the backend dashboard.
GitHub OAuth requires registering an OAuth application and configuring its credentials in the backend. No credentials are bundled. One shared backend can authenticate both sites, but browser login state is separate per origin; seamless cross-domain SSO is not claimed in this local build.