vLLM v0.27.0 Released: What's New, Breaking Changes & Upgrade Guide

What You'll Learn

• What the vLLM v0.27.0 highlights mean for production LLM deployments
• How Kimi K3 and the new Qwen3.5 models run on vLLM out of the box
• Why the PyTorch 2.13 upgrade is a breaking environment change
• The exact upgrade commands and the caveats to watch before you update

The vLLM v0.27.0 release is the biggest inference-engine update of August 2026, and it changes the upgrade math for every team that self-hosts LLMs. Released on August 10, 2026 by the vLLM project on GitHub, it packs 561 commits from 242 contributors, 64 of them new, and it was followed within a day by the v0.27.1 patch release on August 11. The headline is day-one Kimi K3 serving, but the release that matters most for operators is the mandatory PyTorch 2.13.0 move, which the vLLM team flags as a breaking environment change. This guide walks through every major feature, the performance numbers that matter, and the exact steps to upgrade without breaking a running cluster. The inference-engine race is moving fast, and with the four AI giants shipping new models every few weeks, serving speed is now a competitive weapon.

vLLM v0.27.0 at a Glance: The Headlines

The vLLM engine is a high-throughput, memory-efficient inference and serving engine for large language models, originally developed in the Sky Computing Lab at UC Berkeley and now maintained by a community of more than 2,000 contributors. The v0.27.0 release notes, published on the official vLLM GitHub releases page on August 10, 2026, list five headline areas: Kimi K3 support with a full stack landing in one release, the PyTorch 2.13.0 upgrade, FlashAttention 4 on SM100, DeepSeek-V4 performance optimizations, and new model and hardware enablement. The version also lands a Rust frontend gRPC control plane with engine-aware health reporting, abort control, and server and model discovery, which lays the groundwork for production observability.

For teams still on v0.26.x, this is a two-version jump. The move from v0.26.0 to v0.27.0 pulls in PyTorch 2.13.0, torchvision 0.28.0, and Triton 3.7.1, so a plain pip upgrade can resolve to incompatible wheels if your environment pins older torch builds. Plan the upgrade as a coordinated environment change, not a one-line pip bump. If you are tracking the wider framework-release cadence, our langchain-openai 1.4.3 release summary covers the parallel update on the LangChain side of the stack.

Kimi K3 Support: The Full Stack Lands in vLLM v0.27.0

The centerpiece of v0.27.0 is full-stack support for Kimi K3, the 2.8-trillion-parameter Mixture-of-Experts model from Moonshot AI. Kimi K3 activates 16 of its 896 experts per token and is built on the Kimi Delta Attention (KDA) architecture, which stresses an inference stack in ways a dense model cannot. The v0.27.0 release lands core model files and kernels, Python and Rust frontend support, AttnRes kernels, DeepGEMM support, compressed-tensors quantized checkpoints, DSpark AR fusion, and optional shared-expert sharding in a single release. In the official vLLM blog post from July 27, 2026, the team reports serving Kimi K3 at 118 tokens per second without speculative decoding and 370 tokens per second with DSpark, a 3.14x improvement, on 16 NVIDIA GB300 NVL72 GPUs.

That performance matters because Kimi K3 is Moonshot AI's flagship open-weight model, the successor to the K2 line. The v0.27.0 recipe and model registry now list moonshotai/Kimi-K3 as a supported architecture, so the usual vllm serve command works without custom kernels. Teams already running Moonshot models can compare this against the Kimi K2.7 Code launch we covered earlier, and the v0.27.1 patch adds support for quantized DSpark Markov heads, closing the loop for FP8-style K3 deployments.

PyTorch 2.13 Upgrade: The Breaking Change to Plan For

The vLLM v0.27.0 release moves the entire stack to PyTorch 2.13.0, torchvision 0.28.0, and Triton 3.7.1. The release notes mark this explicitly as a breaking environment change, and both the XPU and CPU backends followed the same migration in this release. For GPU operators, the practical impact is that vLLM v0.27.0 wheels expect a torch 2.13 runtime, so any prebuilt kernel cache, custom CUDA extension, or container image built against torch 2.11 or earlier needs a rebuild. The vLLM team also fixed CUDA architecture detection that could produce kernel-less builds on SM121 GPUs, and bumped DeepGEMM from 2.5.0 to 2.6.1 for the MoE kernel stack.

The v0.27.0 release also expanded the Model Runner V2 path, which now covers encoder-only attention, sequence pooling for embedding and classification tasks, encoder token classification, token embedding, BGE-M3 pooling, multimodal models on CPU, and a multi-layer MTP speculator. Performance-critical users should validate their workload against Model Runner V2 before upgrading, because the engine switches to it when the model and pooling configuration qualify. A smooth upgrade here depends on testing the new torch stack against your model set first, which is exactly the kind of migration checklist we outlined for the langchain-openai 1.4.3 update.

FlashAttention 4 on SM100 and Faster DeepSeek-V4 Serving

For NVIDIA Blackwell Ultra users, v0.27.0 integrates FlashAttention 4 on SM100 with FP8 KV cache support and headdim-256 support. FP8 KV cache halves the memory footprint of the attention state, which translates into higher concurrency or longer contexts on the same GPU. The release also adds a JIT warmup infrastructure with runner-owned Triton kernel warmup, which removes the first-request compilation stall that has historically caused latency spikes on cold starts. Combined with the earlier embedding-model work we compared across providers, vLLM now covers a wider slice of the serving stack.

DeepSeek-V4 serving gets a long list of targeted optimizations in v0.27.0: sequence parallelism, a roughly 2x kernel improvement by skipping empty c128 launches, a 3.4 percent end-to-end TTFT reduction by skipping unneeded topk and router work, a 3.9 percent TTFT gain from workspace reuse, a 1.88x kernel speedup by removing a redundant full kernel, adaptive topk width for another 1.0 percent end-to-end gain, and 448 MiB of GPU memory saved in the pipeline-parallel buffer. The team also added a compact MXFP4 indexer KV cache and removed sparse-MLA query-head padding on FlashInfer 0.6.14 and newer. For cost planning on the model itself, our DeepSeek V4 Pro pricing breakdown is a useful companion to these serving improvements.

New Models and Hardware Support in v0.27.0

Beyond Kimi K3, v0.27.0 adds Qwen3.5 support for both text-only dense and MoE variants, K-EXAONE-2.0-750B-A37B from LG AI Research, VaultGemma through the Transformers backend, and the jina-embeddings-v5-text-nano embedding model. On the hardware side, the release adds an sm_107 target for NVIDIA Rubin GPUs with NVLink all-reduce on SM107, plus ROCm support for gfx1250. The sm_107 enablement follows the tracking issue filed in July 2026 and gives early Rubin adopters a build target before the hardware ships broadly.

This hardware breadth is exactly where vLLM differentiates itself from rivals: vLLM runs on NVIDIA, AMD, and TPU-style platforms, while TensorRT-LLM is NVIDIA-only and SGLang covers NVIDIA and AMD. The resilient large-scale serving work in v0.27.0 also matters for fleets: a simplified fault-tolerance framework for data-parallel and expert-parallel deployments behind external load balancers, plus asynchronous preparation for elastic expert-parallel scaling, meaning a failed rank can be evicted and replaced without a full redeploy.

How to Upgrade to vLLM v0.27.0

The official installation path is a clean environment with uv or pip. The current stable release on PyPI is vLLM v0.27.1, published August 11, 2026, which you get with either uv pip install vllm or pip install vllm. The vLLM documentation recommends installing into a fresh conda or venv environment rather than upgrading in place, and for source builds, the precompiled wheel path with VLLM_USE_PRECOMPILED=1 avoids long local compilations.

Before you upgrade, check three things. First, your torch pin: v0.27.0 requires PyTorch 2.13.0, torchvision 0.28.0, and Triton 3.7.1, so an old environment constraint will fail the resolution. Second, your custom kernels and quantization paths: DeepGEMM moved to 2.6.1 and the CUDA architecture detection fix means SM121 builds should be validated again. Third, your model registry: if you run embedding or pooling models, confirm they take the Model Runner V2 path, since that switch can change memory behavior. If you hit the known DeepSeek-V4-Flash error after upgrading from 0.26.0, the fix is to rebuild against the new torch stack rather than reverting.

Conclusion

The vLLM v0.27.0 update is a release you should adopt deliberately. The Kimi K3 full-stack support and the DeepSeek-V4 serving gains are strong reasons to move, but the PyTorch 2.13.0 breaking change means the upgrade is a planned environment migration, not a background pip bump. Start by validating your model set on a staging node with the new torch stack, then roll out v0.27.1, the current stable patch, to production. The vLLM team describes v0.27.0 as its largest community release of the season, and with FlashAttention 4 on SM100 and early Rubin enablement, the release sets the baseline for the rest of 2026. As the team put it in the official release notes, the goal is faster, more resilient serving for every open-weight model that ships this quarter.