DFlash 2
#4 in KI-Inferenz-HardwareUnknown · v2 · 4× · tolest 20. Aug. 2026
54
Momentum
DFlash 2 is a "drafter" for speculative decoding of large language models, released by Inco AI (building on the original DFlash technique developed at Z Lab). It is not physical hardware, but an Open Source model/software method that predicts the output of a target LLM (e.g., Qwen3.8-27B, Meta Muse Glimmer) in parallel blocks, thereby increasing Inference speed without changing output quality. The models are available open source under Apache 2.0 license on Hugging Face and run on common Inference engines such as SGLang, vLLM, llama.cpp, and oMLX/MLX on Apple Silicon.
Momentum-Verloop
22.05.20.08.
Features
| License | Apache 2.0 |
| Platform | Runs in SGLang, vLLM, llama.cpp, and oMLX/MLX (Apple Silicon) from day one |
| Price | Free (open-source model, no purchase price) |
| Release Date | August 18, 2026 |
| Memory | Model size 2B parameters (BF16 ~3.86 GB; Q4_K_M ~1.14 GB; Q8_0 ~2.06 GB) |
| Availability | Publicly available on Hugging Face (incoai/z-lab collections) and GitHub (z-lab/dflash) |