z-lab released gemma-4-26B-A4B-it-DFlash. Anybody tried it yet?

On May 8th, a lab called z-lab posted to LocalLLaMA with a release announcement and an open question: has anyone tried it yet? The model in question is gemma-4-26B-A4B-it-DFlash, and the name alone tells a story. Gemma 4 as the architectural foundation. Twenty-six billion total parameters with an A4B tag that implies only around 4 billion of those are active at any given time — the fingerprint of a mixture-of-experts or sparse activation design, where the model carries weight it doesn't always spend. Then "DFlash," which likely signals either a distillation variant, a flash attention implementation, or both layered together. The post scored 4.3 on LocalLLaMA, which is modest — enough to surface, not enough to cause a stir. The forum's response so far appears to be the same question the original poster asked. Nobody has tried it yet, or nobody who has tried it has said so publicly.

If confirmed, here is what this means. A 26B total parameter model running on roughly 4B active parameters is a genuinely interesting efficiency target — you get a model that behaves like something larger without paying the full inference cost at every token. Built on Gemma 4, it inherits whatever architectural improvements Google baked into that generation, then adds whatever z-lab has done on top with the distillation or flash attention layer. For local runners, that active parameter count matters enormously: it determines what hardware this actually fits on comfortably. If the efficiency claims hold, this could run well on consumer hardware that would struggle with a dense model of comparable capability. The broader implication is that small, unnamed labs are continuing to compress capable models into deployable shapes faster than the community can track them, which is either exciting or quietly alarming depending on how much you care about the provenance of the models you run.

Watch for independent benchmark results or any user reports confirming inference behaviour on consumer hardware — and look for whether z-lab surfaces any technical documentation that clarifies what "DFlash" actually means in their implementation.

NewsHive monitors these sources continuously. All signal titles above link to the original reporting.

Intelligence by NewsHive. Need help navigating what this means for your business? Contact GeekyBee →