Local Model Desk

简体中文 · Windows downloads · License

Local Model Desk is a Windows launcher for dual-GPU Qwen3.8-Flash-Next on Strata. It starts and unloads models, keeps the selected OpenCode model and context in sync, embeds Strata telemetry, and provides per-card NVIDIA power controls.

Workbench shown with a simulated local service.

Install and start

1. Use the Windows x64 portable ZIP for this build and check it against the accompanying SHA-256 file. GitHub's Source code links contain repository snapshots.
2. Extract the entire archive and run LocalModelDesk.exe. Keep its _internal directory alongside the executable. The ZIP includes the launcher's runtime, source, and documentation.
3. In Settings, select the existing Strata deployment root. The launcher targets terminal/0.1.57/bin/strata.cmd and the model data folder models/. It finds OpenCode Desktop in the deployment or in the Windows installation. Enter the exact executable in Settings to override discovery. Strata, OpenCode, drivers and model weights are installed separately.
4. Choose IQ2_XS or IQ3_XXS and a 32K–256K context. Check download previews downloads and resource guidance. Start model and open OpenCode starts the local service and backs up the OpenCode configuration before updating it. GPU 0 is the main card and GPU 1 is the helper.
5. To upgrade, close the old launcher and extract the new ZIP into a separate folder. Local Model Desk keeps its preferences in %LOCALAPPDATA%\QwenDesk\Python\settings.json. Save the interface language in Settings and reopen the app to apply it.

OpenCode Desktop 2

OpenCode Desktop 2.2.6 or later uses a separate configuration and session directory under agent/opencode-v2/. Local Model Desk opens the installed desktop directly, configures the selected Strata model, and loads its bundled strata-charts GUI plugin. If an OpenCode background service is already using another configuration, finish its active tasks and accept the restart prompt. On OpenCode 1.x the existing deployment launcher and configuration are still used.

Ask for an interactive chart in chat. The plugin renders a completed strata-chart block as a bar chart with a switch for each supplied metric. For example, ask it to chart two sample decode rates and VRAM values in strata-chart format. The plugin can be inspected under Settings → Extensions → Plugins in OpenCode Desktop.

Context budgets

Strata context	OpenCode input budget	Output cap	Default compaction reserve	Default compaction threshold
32K	28K	4K	2K	26K
64K	56K	8K	4K	52K
128K	112K	16K	8K	104K
256K	224K	32K	16K	208K

K means 1,024 tokens. The threshold assumes automatic compaction is enabled and uses the default reserve. The output cap includes reasoning tokens. Available system memory and VRAM affect how much of the configured context you can use comfortably.

Images, monitoring, and power

Enable image input on Home, check the download plan, and start or switch the model. First-time setup may download a vision encoder of about 0.9 GB. Open a new OpenCode chat and attach an image. The image tokens share the chosen context window with text. If attachment delivery fails, try the same picture in Strata's browser chat at http://127.0.0.1:8080.

The Monitor page shows prompt progress, output rate, GPU usage, and recent requests. GPU Power queries each card via the installed NVIDIA driver. Enter watts within its reported range, click Apply, and approve the Windows prompt when shown. The page reads the limit back after applying it. Restore default uses the driver's default for that card. Check the current limit again after a reboot or driver reload.

Troubleshoot

• For missing Strata or OpenCode files, check the configured deployment root and paths above.
• The launcher uses port 8080. If a different service occupies it, stop that service from its original launcher.
• Check Logs for download and startup errors. On the same PC, query http://127.0.0.1:8080/health to check model, max_context, loaded and images.
• A vision setup can reuse the existing text weights. If the image state is off after launch, check the Strata log for the vision encoder.
• For power changes, check the driver's reported range and the administrator prompt. Driver and hardware support vary.

Screenshots

Embedded Strata monitor

Prompt progress, recent requests, and dual-GPU telemetry. The readings shown are simulated.

GPU power limits

Per-card power draw, allowed limits, and restore controls. The readings shown are simulated.

Settings

Strata installation path, appearance, and interface language.

Build and test from source

Use Python 3.12 on Windows x64. Run these commands in PowerShell from the repository root.

py -3.12 -m venv .venv
& .\.venv\Scripts\python.exe -m pip install -r .\source\requirements.txt
& .\.venv\Scripts\python.exe .\source\app.py
& .\.venv\Scripts\python.exe -m unittest discover -s .\source
& .\.venv\Scripts\python.exe .\source\app.py --smoke-test .\validation
& .\source\Build.ps1 -Python .\.venv\Scripts\python.exe

The portable build is written to dist/LocalModelDesk/. The smoke test uses simulated GPU writes and service state.

License

The application source is available under the MIT license. The embedded Overview renderer retains the original Strata notice. See the third-party notices for the portable distribution.