Keras 3.15 Ships Native vLLM Integration, Pulling Production Inference Into the Keras Ecosystem
Keras 3.15 adds built-in vLLM serving and speculative decoding, tightening its grip on the full model-to-production pipeline.
3. Keras 3.15 Ships Native vLLM Integration, Pulling Production Inference Into the Keras Ecosystem
François Chollet announced Keras 3.15 on August 7, 2026, with three headline additions: native vLLM integration for serving KerasHub models, built-in speculative decoding across all KerasHub CausalLMs, and new model architectures in KerasHub including Gemma 4 variants compatible with all HuggingFace checkpoints for those architectures. Chollet described the vLLM integration as delivering "large performance gains." A community meeting is scheduled for Friday, August 9 at 10am PT, where the team will present the latest ecosystem developments.
The strategic move here is about closing the gap between training and serving inside a single framework. Until now, teams using Keras for model development still had to reach outside the ecosystem to wire up production-grade inference, typically through vLLM or TGI directly. That hand-off created friction and gave frameworks like PyTorch-native tooling a structural advantage at the deployment stage. By absorbing vLLM as a first-class integration, Keras now competes across the full workflow. The HuggingFace checkpoint compatibility for Gemma 4 variants is equally pointed: it lets Keras users pull weights from the dominant model-sharing platform without leaving the Keras abstraction layer, directly contesting HuggingFace's position as the connective tissue of the open-source model stack.
Speculative decoding as a default capability for all CausalLMs is the quieter but durable part of this release. It shifts inference optimization from an expert-configured add-on to a baseline expectation. Watch whether PyTorch-native serving stacks respond by tightening their own framework integrations, and whether the upcoming community meeting surfaces details on the "pluggable backends" topic Chollet flagged, which could signal the next layer of the Keras infrastructure push.
Source: @fchollet on X