What changed in vllm, release by release
The lines under each release are vllm's own release-note bullets with pull request numbers, names and links removed, grouped by its own headings. Wording is theirs. Not affiliated with vllm.
v0.34.0
- Changed10 new models (Mistral Large 4, MiniMax-M3.1-Flash, Nemotron-3.5-Transcribe, and more!)
- ChangedExtensible KV cache is on by default: vLLM sizes the KV cache from memory actually free after warmup, so most models get a bigger cache and more…
v0.33.0rc3
- BreakingBreaking changes: Medusa and the MLP speculator removed ; the legacy Ray compiled-graph executor removed ; --mamba-cache-mode deprecated ; requests…
- NewNew models: Reflection Beam-1.0 (501B MoE, 23B active) with its DFlash2 drafter , Yandex AliceAI-Foundation-80B-A3B , Recursive Flint (40B MoE, 2.5B…
- NewCUDA 12.9 dropped: release wheels and images are CUDA 13.0 only. The CUDA 12.9 wheels and the -cu129 / -cu129-ubuntu2404 images are no longer…
- ChangedSource distribution tarball
- ChangedCPU Python wheels for
x86_64, arm64, and macOS
v0.32.0
- BreakingBreaking changes: Python 3.10 dropped ; default CORS origins restricted ; transformers=5.16.1,5.20.0 and tokenizers=0.23.2 ; torch.compile sequence…
- NewStructured decisions and generic parsers: the new POST /v1/systemone endpoint answers choice, noul and score questions over a structured state , is…
- NewLow-precision KV cache everywhere: NVFP4 KV cache now runs on SM8x and SM12x , SM90 and with XQA decode on SM12x , and on MiniMax-M3 sparse attention…
- ChangedSource distribution tarball
- ChangedCPU Python wheels for
x86_64, arm64, and macOS
v0.31.0
- BreakingBreaking changes: per-request multimodal kwargs gated ; tokenizer_mode="slow" removed ; --enable-mamba-fine-grained-prefix-cache renamed to…
- NewDeepSeek-V4.1-Flash performance: FlashMLA mega attention with the V4.1 NVFP4 compressed KV cache is now the SM100 default ; DeepGEMM sparse MQA…
- NewFast restart: the new vllm preload CLI launches the weight-cache daemon that keeps post-quantized weights resident in GPU memory across engine…
- ChangedSource distribution tarball
- ChangedCPU Python wheels for
x86_64, arm64, and macOS
v0.30.0
- BreakingBreaking changes: scale-out endpoints are opt-in on plain vllm serve via --enable-scale-out, replacing VLLM_ENABLE_SCALE_OUT_ENDPOINTS ; GPTQ…
- NewNew models: DeepSeek-V4.1-Flash with the whole KV stored in MXFP8 through the FlashMLA V4.1 record on SM100 , DeepGEMM Mega-mHC , and async Engram…
- NewFast Start: a persistent per-GPU weight-cache daemon holds post-quantized, TP-sharded weights in GPU memory so restarting engines map them over CUDA…
- ChangedSource distribution tarball
- ChangedCUDA 13.0 Python whee
v0.29.0
- BreakingBreaking changes: ten deprecated model architectures removed ; FlexOlmo, Olmo3 and Hunyuan V1/VL migrated to the Transformers modeling backend ; PyAV…
- NewModel Runner V2 is now the default for all models , completing the rollout that began with pooling models . MRV2 also gained CUDA graph memory…
- NewNew models: Hy4-preview, Tencent's 770B/49B-active MoE with Gated DeepSeek Sparse Attention and native MTP ; Qwen3.8-Flash-Next with BF16/FP8/NVFP4…
- ChangedSource distribution tarball
- ChangedCPU Python wheels for
x86_64, arm64, and macOS
v0.28.0
- BreakingBreaking changes: bitsandbytes support migrated to an out-of-tree plugin ; Transformers bumped to 5.15.0 ; the deprecated
calculate_kv_scalesruntime… - NewKimi-K3 performance push: a major optimization effort for Kimi-K3 across the stack — Decode Context Parallel (DCP) support , fused FlashKDA decode…
- NewDeepSeek V4: sparse MLA now works end-to-end for plain decode, MTP, and DSpark speculative decoding , joined by AMD Quark NVFP4 support ,…
- ChangedSource distribution tarball
- ChangedCPU Python wheels for
x86_64, arm64, and macOS
v0.27.1
- ChangedSupport quantized DSpark Markov heads
v0.27.0
- NewKimi K3 support with a full stack landing in one release: core model files and kernels , Python and Rust frontends, AttnRes kernels , DeepGEMM…
- NewMore new models: Qwen3.5 text-only dense and MoE models with EVS video token pruning , K-EXAONE-2.0-750B-A37B , VaultGemma via the Transformers…
- ChangedKimi K3: new model with model files and kernels , Python frontend , Rust frontend , AttnRes kernels , DeepGEMM support , DSpark AR fusion , and…
- ChangedNew models: Qwen3.5 text-only dense and MoE , K-EXAONE-2.0-750B-A37B , VaultGemma via Transformers backend , jina-embeddings-v5-text-nano with…
v0.26.0
- NewNew Inkling model family with a full support stack: base modeling , piecewise CUDA graph support , Hopper FA4 relative attention , MTP=1 speculative…
- NewDeepSeek-V4 performance push across vendors: a specialized routing kernel (2.94% E2E TPOT, ),
fused_topk_bias(1.5–2x kernel, ), and redundant… - ChangedNew models: Inkling family , BertForMaskedLM , RobertaForTokenClassification / XLMRobertaForTokenClassification , LongCat-Flash-Lite n-gram embedding…
- ChangedTransformers backend migrations: Olmo/Olmo2 , MistralLarge3 to AutoWeightsLoader , HunyuanVL native transformers processor for transformers 5.13 .
How we calculate
- Days between releases: The median of the gaps, in days, between the release dates listed above.
Data: vllm-project/vllm's GitHub releases page, read November 18, 2026. Every other figure on this page is copied from the record, not calculated by us.
Get this for your repo, every release
vllm publishes release notes that people can read. FoxPlug writes yours from what you ship: a changelog entry and posts for the places your users are, each approved by you.
Get this for your repo, every releaseTry it on your repo nowYour changelog, written for you
FoxPlug writes your changelog from your releases and commits in plain English and publishes it as a page and a feed your users can follow. The page above shows what that looks like for a product people really use.
Write my changelog See the changelog FoxPlug wrote from its own commits
Free to start. Pro is a month.
FoxPlug Pro
- A public changelog page and feed for each project
- Written from your releases and commits, in plain English your users read
- Connect GitHub and new releases become new entries