vLLM v0.26.0 Released July 27, 2026 with Inkling Support and KV Offloading Maturity
vLLM v0.26.0 was released on July 27, 2026 with 411 commits from 212 contributors (61 new), introducing full Inkling model family support, matured KV offloadin…
Published on MyPrivateClaw
Jul 26, 2026, 3:24 PM UTC
Coverage date
Jul 25, 2026
Last updated
Jul 29, 2026, 3:03 AM UTC
News summary
vLLM v0.26.0 was released on July 27, 2026 with 411 commits from 212 contributors (61 new). The release adds full Inkling model family support including base modeling, piecewise CUDA graph support, Hopper FA4 relative attention, MTP=1 speculative decoding, LoRA, and standard ModelOpt NVFP4 quantization. vLLM v0.26.0 introduces fp32 lm head for generation models via head dtype, extended to the LoRA path and with a ROCm torch.mm fast path. The release updates to Transformers 5.13.0 and migrates Olmo/Olmo2, MistralLarge3, and HunyuanVL models to the Transformers modeling backend. vLLM v0.26.0 Rust frontend gained multimodal video and audio support, a Seed OSS tool parser, and a native vllm bench port. vLLM v0.26.0 substantially matured KV offloading and tiered secondary storage with object store secondary tiers, DP replica aware tiering, encoder cache connectors including CPU offloading, a…