DTree on MLX ... tiny win over DFlash on Qwen3.5-4B (M2)..

On April 15th, a LocalLLaMA user published benchmark results from an M2 machine running Qwen3.5-4B inference, comparing two attention implementations on Apple's MLX framework: DFlash, the flash attention approach that has become the quiet default for efficient computation on Apple Silicon, and DTree, a decision-tree-based attention mechanism that most practitioners haven't thought much about. DTree won. Barely — the poster themselves called it a "tiny" margin — but it won. That's the thing worth sitting with. Flash attention didn't become the default by accident; it earned that position through a string of benchmark victories and strong theoretical grounding. When something beats it, even by a whisker, on a specific hardware-model pairing, it raises a question that's easy to ignore and harder to answer: is this an outlier, or is it the first sign that different architectures reward different attention strategies in ways we haven't fully mapped yet?

If confirmed, here is what this means. The assumption that flash attention is the correct default for MLX inference on Apple Silicon deserves re-examination — at least for mid-size models in the 4B parameter range. For developers building local inference pipelines on M-series chips, the practical implication is simple: benchmark both, because your intuition about which to reach for may cost you performance. The deeper implication is more interesting. If decision-tree attention can outperform flash attention on specific model-hardware combinations, then the optimization landscape for edge inference is messier and more architecture-dependent than the field's current consensus suggests. That matters for anyone building tooling, defaults, or recommendations on top of MLX — which increasingly means a significant slice of the local AI development community.

Watch for independent reproductions on other M-series chips and with different model sizes, and for any response from the MLX development community that either validates or contextualizes the result.

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 →