News
vLLM 0.26 expands model and quantization support
The v0.26.0 release adds an Inkling model support stack, Hopper attention work, speculative decoding support, LoRA, and NVFP4 quantization.
vLLMInferenceOpen Source
vLLM 0.26.0 contains 411 commits from 212 contributors. Its highlighted work includes a complete support stack for the Inkling model family, piecewise CUDA graphs, Hopper FA4 relative attention, MTP speculative decoding, LoRA, and ModelOpt NVFP4 quantization.
Why it matters: The release demonstrates how quickly inference runtimes must evolve to support new model architectures and hardware-specific execution paths. Teams should review compatibility and benchmark their own serving workloads before upgrading.
Source attribution
Original reporting and release information: vLLM GitHub Releases.