DFlash Console

Local Windows control panel for DFlash speculative-decoding stacks and a unified model runtime. Load GGUF, Hugging Face, speech, and embedding models from one UI, then talk to them through a single OpenAI-compatible port.

Status: Public preview for local, single-user Windows use.

Developer: ILAN AVIV · UI: http://127.0.0.1:8900/ · Version: v0.3.208

Download (Windows)

Installed desktop apps check for updates from the latest GitHub Release (latest.json + the setup EXE). Windows may show an Unknown publisher warning because the installer is unsigned.

Engines

Pick an engine in Engines, Models, or Playground before you load.

Install the optional engines from Settings → Downloads & engines or the first-run wizard. They stay out of the Windows installer so the download stays small.

DFlash 1 and DFlash 2

A DFlash stack is a target GGUF plus a smaller draft / accelerator used for speculative decoding.

The catalog can filter DFlash and DFlash 2 separately. On a target model, right-click Find and attach draft — the app searches the local library and Hugging Face, checks architecture compatibility, then registers and attaches the matching draft.

What else it does

Quick start

pip install dflash-console
dflash serve

Or install the Windows EXE and open the app. The UI is http://127.0.0.1:8900/.

Typical first session:

1. Open Engines, turn Running on, and pick DFlash, vLLM, Transformers, or FreeToken.
2. Load a model from the dropdown or the Models tab — a loading card appears immediately.
3. For a DFlash GGUF, right-click and Find and attach draft if you want speculative decoding.
4. Chat in the Playground, or point any OpenAI client at http://127.0.0.1:8001/v1 with optional X-DFlash-Client: YourApp so Engines shows who is using each model.

Recent improvements (v0.3.208)

• Engine standby — Running toggle gates load/chat until you arm the pipeline
• External GPU cards — OneVoice, LM Studio, and other apps on the GPU; compact mobile layout; loading state expires when models are ready
• Other GPU processes — per-process VRAM chips for non-model apps
• Unload fix — engine card Unload works reliably while the list refreshes
• Client identity — X-DFlash-Client header shows Active client on each card
• VRAM preflight — load plans suggest which engine to unload first; co-resident 12B+31B pairs on one GPU
• HF catalog — local index for instant Model catalog search
• GPU performance mode — Balanced / Performance / Inference / Power in Settings → Hardware

From a git checkout: copy config.example.json to config.json, then .\server.ps1. Full walkthrough: docs/USER-GUIDE.md.

Terminal CLI

dflash serve
dflash list
dflash list --dflash
dflash list --vllm
dflash list --transformers
dflash load qwen
dflash chat "hello"
dflash search "qwen dflash2"

The Console must be running except for help, version, serve, and install. Full command list: docs/CLI.md.

OpenAI gateway and API

Point clients at:

http://127.0.0.1:8001/v1

The gateway routes chat, embeddings, TTS, and STT to the loaded engine. Model names are tolerant (engine id, file name, or an alias such as gpt-4o).

Client identity: send X-DFlash-Client: YourApp on load and chat requests so the Engines page shows the Active client for each model. The gateway forwards this header. The banner updates on every chat or embed call while the model stays loaded. Without the header, callers appear as Unknown API client. See docs/CLIENT-IDENTITY.md.

Selected Console routes (UI/API on port 8900):

Load a Hugging Face folder on a specific engine:

curl -X POST http://127.0.0.1:8900/api/models/load \
  -H "Content-Type: application/json" \
  -H "X-DFlash-Client: MyApp" \
  -d "{\"path\": \"C:\\\\models\\\\org\\\\model\", \"runtime_id\": \"vllm\"}"

Use "runtime_id": "transformers" or "freetoken" the same way.

Full list: Documentation in the app, or Swagger at http://127.0.0.1:8900/docs.

Requirements

• Windows 10+
• Python 3.10+
• PowerShell 7+ (pwsh)
• NVIDIA GPU recommended for multi-model loads
• Optional: WSL2 Ubuntu + CUDA for FreeToken
• Optional: Node.js 22.12+ only if you build the Electron shell from source

Configuration and security

Copy config.example.json → config.json (never commit the live file).

The Console binds to loopback. It is for one trusted user on one PC. Do not expose ports 8900 or 8001 to a LAN or the internet without adding your own auth. Keep Hugging Face tokens in the environment (HF_TOKEN), not in config.json.

Community

• Questions: Discussions
• Bugs: Issues
• Security: SECURITY.md

Include the version from About. Do not paste config.json, tokens, or model weights.

License

GNU AGPL v3 or later. See LICENSE, NOTICE.md, and TRADEMARKS.md.

DFlash Console is developed by ILAN AVIV. Related project: DFlash.