Z‑lab Releases DFlash‑Enabled Gemma 4 26B A4B‑it Model
Z‑lab has released the DFlash‑enabled Gemma 4 26B A4B‑it model, offering faster speculative decoding and stateful KV cache for longer sessions.
Integrate the new Gemma‑4‑26B‑A4B‑it‑DFlash model into vLLM and benchmark against the base model to assess speed gains.
Summary
Z‑lab has released a new variant of the Gemma 4 26B model that incorporates DFlash speculative decoding, available as gemma‑4‑26B‑A4B‑it‑DFlash on Hugging Face. The DFlash mechanism provides stateful KV cache and parallel block diffusion drafting, which is expected to improve performance for longer sessions. Unlike MTP, DFlash maintains persistent state across iterations for context buffers, KV cache positions, and RoPE offsets, potentially reducing cache ballooning as context grows.
The model is built on the A4B‑it quantization scheme and is optimized for vLLM inference. Early reports suggest significant speed gains over the base Gemma 4 26B model, especially when used with large context sizes. The release also includes documentation on how to enable DFlash in vLLM and notes that support for LCPP is still pending. Users are encouraged to benchmark the new model against the original to quantify performance improvements.
Key changes
- New Gemma 4 26B A4B‑it model with DFlash speculative decoding.
- Stateful KV cache and parallel block diffusion drafting for long sessions.
- Optimized for vLLM inference with A4B‑it quantization.
- Available on Hugging Face as gemma‑4‑26B‑A4B‑it‑DFlash.
- Documentation includes DFlash enablement steps for vLLM.
- LCPP support still pending.