I built DFlash Console, a local Windows control panel for running GGUF and Hugging Face models through one UI and one OpenAI-compatible port.
The problem I kept hitting: a target GGUF, a smaller DFlash draft for speculative decoding, llama-server flags, and a separate path if I wanted vLLM or Transformers. I wanted one library that can find a compatible draft and attach it.
Engines (you pick one before load):
- DFlash / llama-server — GGUF chat, vision, OCR, embeddings. Default local path. DFlash 1 drafts work with the bundled llama-server. DFlash 2 (Gemma 4 / Qwen 3.8-style drafts) needs a llama.cpp build that actually has DFlash 2.
- vLLM — Hugging Face SafeTensors on NVIDIA. Optional, installed on demand.
- Transformers — HF SafeTensors on more PCs (CPU or GPU). Optional.
- FreeToken — large HF MoE folders via WSL2 + CUDA. Optional.
On a target GGUF you can right-click Find and attach draft. It searches the local library and Hugging Face, checks architecture compatibility, then registers and attaches the matching DFlash 1 or DFlash 2 draft.
Also: catalog search/download from Hugging Face into your model folders; Playground for chat / speak / transcribe / embed; Piper TTS and whisper.cpp STT; gateway at 127.0.0.1:8001/v1. The UI binds to loopback (127.0.0.1:8900). Single-user, not a LAN server.
Install (Windows):
pip install dflash-console
dflash serve
Repo (AGPL-3.0, public preview): GitHub - ilan4ever/Dflash-Console: Local Windows control panel for llama-server and DFlash speculative-decoding engines. · GitHub
PyPI: Client Challenge
Related upstream: z-lab/dflash on GitHub (DFlash 1 and DFlash 2 drafts). There is also a Windows setup/portable build on GitHub Releases. It is unsigned, so SmartScreen may warn. I would start with pip.
I would like feedback on: DFlash 2 llama.cpp builds that actually work on Windows, draft matching edge cases, and whether the HF catalog → local library flow is the right shape. Happy to answer technical questions here.