Loading documentation…
Pick an Engine (DFlash, vLLM, or Transformers), then a model, then Load.
Use vLLM for Hugging Face SafeTensors models. GGUF files stay on DFlash.
| Model | Family | Scale | Source | Disk | Updated | |
|---|---|---|---|---|---|---|
| Loading your model library… Scanning configured folders. Your models will appear here shortly. | ||||||
Models downloading from Hugging Face appear here with live progress.
Connect other DFlash Console instances on your network. Check health and send a test chat through each node.
Loading nodes…
Loading documentation…
A focused control panel for llama-server engines, DFlash speculative decoding, local model libraries, and live runtime diagnostics.
DFlash Console starts and monitors local engine profiles, loads and unloads GGUF models, tracks GPU and token statistics, searches Hugging Face, and exposes OpenAI-compatible local endpoints.
The Console is designed for one trusted user on one machine. It binds to loopback by default and does not provide multi-user authentication. Do not expose it beyond loopback without adding an authenticated access layer.
DFlash-Console-Setup-*-x64.exe)pip install dflash-console then dflash serveconfig.example.json to config.json, run .\run.ps1Public preview. Report bugs on GitHub Issues. Only one Console API should run on port 8900 at a time.
The Windows installer is a thin Electron shell. The backend, documentation, model files, logs, and native llama-server binaries remain in the selected Console data root so large runtime assets are not copied into every installation.
Configure DFLASH_CONSOLE_ROOT or choose the data
root when the packaged app first starts.
The GNU AGPL requires retaining the copyright and license notices and providing corresponding source when its terms apply. When you redistribute or build on DFlash Console, please also link to the DFlash Console source repository in your README, About page, or project documentation. The DFlash name and logo are governed separately by TRADEMARKS.md.
DFlash Console is free and open source, developed in spare time on local hardware. If it saves you time or helps your projects, a small donation supports development, hardware, and hosting.
Donations go through PayPal and are not required to use any feature of DFlash Console.
Select a model to view details, README, and download GGUF files.
Everything you download after setup: LLM engines (vLLM, Transformers), speech runtimes, and Hugging Face model downloads. Install an engine here, then pick it on the Engines, Playground, or Models tab and load a model.
Extra downloads you install after setup appear first. Speech and GGUF engines bundled with the app are listed below.
Loading components…
Hugging Face downloads currently in progress
None in progress.
New Console releases may ship updated speech bundles. Check for signed Windows updates in Desktop app settings.
Tray, startup, and other preferences for the Windows installer. The browser UI ignores these options.
The browser and the desktop app must show the same address. Both read one config file from this server.
Open this in a browser to see the same Settings as the desktop app.
—
Shared hardware and engine settings
—
What happens when you close the window or restart your PC.
Open DFlash Console when you sign in.
Minimizing or closing the window hides it in the tray instead of the taskbar.
When Windows starts DFlash at sign-in, keep the window hidden until you open it from the tray. Opening the desktop shortcut always shows the window.
Brief loading window while the local engine starts.
Allow DFlash Console to check periodically. You can always check manually below.
Version and folders used by the desktop shell.
—
Backend, models, and config (server.ps1, api, static).
—
Tray and startup settings for this install.
—
Check the DFlash Console website for signed Windows updates. Downloads are verified before installation.
Updates are available only in the desktop app.
Choose where Hugging Face downloads land and which folders the Models tab scans on this PC.
New Hugging Face downloads save here
Parallel HTTP ranges for Hugging Face downloads. Files under 32 MiB always use one connection.
Downloads a small sample from Hugging Face and compares 1, 2, 4, 6, and 8 parallel connections on this PC.
Enable folders to scan for models on this PC. Use Scan PC to find Piper, Whisper, GGUF, OCR, and other installs automatically.
Read-only folders used by DFlash Console. Export or import config and launch presets when moving to a new machine.
Engine profiles, libraries, and preferences
—
llama-server binaries, draft models, and launch scripts
—
Default folder scanned for GGUF models
—
Boot and load output for each engine profile
—
Generated .ini files before each router launch
—
Address where this app is served
—
Back up or move your config and generated launch presets.
Detecting GPUs…
Used when an engine is set to pick the GPU automatically. A GPU you choose by hand on Engines still wins. Changing this reloads loaded models that need a new layout — you do not restart DFlash Console.
Prefer dedicated GPU memory over shared memory for weights.
Turn off to keep the active KV cache in system RAM. Disk is used only for saved prompt/slot cache, not live inference.
Refreshes every few seconds while settings stay open.
Network, runtime, and launch defaults for each engine in config.json. Pick a profile first — changes apply to that engine only.
Which configured engine you are editing
OpenAI-compatible endpoint for this profile.
API port for this engine
127.0.0.1 keeps the API local-only
—
One friendly OpenAI-compatible URL for any client (chat, embeddings, audio)
Single port like LM Studio's 1234 (default 8001). Restart the app after changing.
Engine that /v1/chat/completions routes to
—
Default limits when this engine loads a model.
Starting context on load — external API requests may grow it up to the limit below
Hard cap — API requests can never exceed this
Minutes before this Console engine unloads its model when idle · 0 = never · does not affect LM Studio, Ollama, or other external apps
DFlash preset and GPU override for this profile.
Target/draft pairing and speculative decoding mode
Override automatic GPU selection
Manage engines, load models, and proxy chat completions.
Full guide: open Documentation in the sidebar, or Swagger UI.
Connect Cursor, Claude Desktop, and other tools to your local engines. MCP hosting is not active yet — use the OpenAI-compatible URLs below.
In-console MCP discovery and registration
These URLs work today with Playground, API clients, and many integrations
Starter config for Cursor, Claude Desktop, or other MCP-aware apps
{}
Device policy, CPU fallback, and memory budget for Piper, Whisper, and other non-llama runtimes. To install vLLM or Transformers, use Downloads & engines.
vLLM and Transformers downloads are in one place — not here.
Device policy, CPU fallback, and memory budget per runtime
Loading…
Which Console runtimes or external apps are holding VRAM right now
Optional guards when VRAM is busy
When GPU contention recommends it, unload other Console engines before loading the target
Show a reminder in the Playground when the active TTS/STT runtime is set to CPU
Writes runtimes[] back to config.json