vLLM 0.28.0 Optimizes Open-Weight Inference

This release makes model architecture inseparable from serving economics: open-weight winners increasingly depend on the stack that turns exotic designs into affordable tokens. The gains remain hardware- and workload-dependent, so teams should benchmark their exact concurrency, context, quantization, and accelerator mix.

• –Kimi K3 gains fused FlashKDA kernels, decode-context parallelism, adaptive speculative decoding, and expert sharding that can save about 17 GiB per GPU.
• –vLLM reports up to 3.14× higher Kimi K3 throughput with DSpark on 16 NVIDIA GB300 GPUs. [Kimi K3 benchmarks](https://vllm-project.github.io/2026/07/27/k3.html)
• –DeepSeek V4 receives end-to-end sparse MLA support across decode, MTP, and DSpark, plus expanded AMD ROCm and NVFP4 support.
• –Qwen3.8 support on AMD ROCm broadens accelerator choice, while tiered KV-cache offloading and E/P/D disaggregation improve large-scale serving flexibility.
• –CUDA 13.0 becomes the default, but bitsandbytes migration and removed deprecated APIs mean production upgrades require compatibility testing.