GoKawiilvLLM 0.28.0 ships major performance upgrades for Kimi-K3 and DeepSeek V4
The vLLM project released version 0.28.0, combining 584 commits from 270 contributors. The update centers on deep performance work for Kimi-K3, including new decode context parallelism, fused kernels, memory-saving expert sharding, and ROCm support, alongside DeepSeek V4 improvements such as end-to-end sparse MLA and AMD Quark NVFP4 support.
GoKawiil's interpretation of the reporting above, not reported fact.
These optimizations translate directly into faster inference and lower memory costs for large mixture-of-experts models, which are increasingly central to production LLM deployments. Expanding ROCm compatibility also signals vLLM's push to support AMD hardware as a viable alternative to NVIDIA for serving these massive models.
- Kimi-K3 gains multiple kernel and parallelism optimizations, including a claimed 60% TTFT improvement and 17 GiB memory savings per GPU.
- DeepSeek V4 now supports sparse MLA across decode and speculative decoding paths, plus AMD Quark NVFP4 quantization.
- Both major models now have expanded ROCm support, broadening vLLM's hardware compatibility beyond NVIDIA GPUs.
NVIDIA GeForce RTX 4090 Founders Edition — vLLM's latest release is all about squeezing more throughput out of GPUs, and the RTX 4090 is the go-to card for running large language models locally. With its huge VRAM pool and tensor cores, it lets you actually put features like fused decode kernels and speculative decoding to work on your own rig.
See NVIDIA GeForce RTX 4090 Founders Edition on Amazon → Affiliate link — we may earn a commission on purchases, at no extra cost to you. Product picked by AI based on this article; it is not a tested recommendation.Source: github.com, 2026-08-29
Published there as: “vLLM v0.28.0”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.