# vLLM v0.33.0rc3 — Release & Upgrade Intelligence
- **Product:** vLLM (https://vllm.ai)
- **Version:** v0.33.0rc3
- **Released:** Nov 02, 2026
- **Impact Score:** 7/10
- **Breaking Changes:** Source flagged
- **Canonical Analysis:** https://changelogs.us/vllm/updates/v0.33.0rc3/

## Executive Summary (TL;DR)
vLLM v0.33.0rc3. Source highlight: This release features 927 commits from 389 contributors (122 new)!

## Breaking Changes & Deprecations
- Breaking changes**: Medusa and the MLP speculator removed (#61701); the legacy Ray compiled-graph executor removed (#61712, #62228); `--mamba-cache-mode` deprecated (#60043); requests sent while generation is paused now get HTTP 503 (#51488); the NIXL connector needs NIXL 1.5 or newer (#61032); LM Format Enforcer and `--ir-op-priority` deprecated (#63423, #63371).
- Breaking Changes & Deprecations * CUDA 12.9 wheels and images are no longer published (#62575). * Medusa and the MLP speculator were removed (#61701). The legacy Ray compiled-graph executor was removed and `VLLM_USE_RAY_V2_EXECUTOR_BACKEND` is gone; `RayExecutorV2` is the only Ray executor (#61712, #62228).

## New Features
- **Other models**: Cohere2MoE attention/MoE dual stream, single TP all-reduce and fused add + RMSNorm (1.14-1.36x TPOT, on by default, #62101); Beam TP all-reduce fused with norms and MoE finalize (#62366); MiniMax-M3 CUTLASS MSA decode crossover (#55838); Granite MoE as an EAGLE3/DFlash/DSpark target (#61110); NemotronHMTP sharing target embeddings (#62194); LoRA for BERT cross-encoders and token classification (#61279, #61391).

## Improvements
- **NVIDIA**: Kimi-K3 and DeepSeek-V4.1-Flash latency-specialized decode on GB300 (#61867, #61868); FA4 sparse MLA decode (#55866); opt-in CuTe DSL MLA context prefill (#62069); DeepGEMM MegaMoE with NVFP4 routed experts (#61538) and warmed-up kernels (#60468, #61657); Triton GDN decode tuned on SM100 (#61355); GLM-5.2/5.3 low-latency BF16 GEMM on SM107 (#62843); pre-SM90 FlashInfer spec decode in full CUDA graphs (73.6 to 108.1 tok/s on SM80, #60532); `cooperative_topk` sized by row length (#6217
- **Intel XPU**: SYCL GDN prefill backend selection (#56835), torch MXFP4 linear backend (#51965), SYCL FlashAttention for prefix-LM text batches (#62014) and NIXL 1.5.0 in the image (#60356).
- **CPU**: zentorch 2.14.0.1 with float16 on Zen CPUs with AVX512-FP16 (#61411); MRV2 Triton warmup at startup removing an 11-20 s first-request compile (#62888); INT8 W8A8 MoE brgemm (#60740); hybrid-model prefix caching (#61805); and Run:ai Model Streamer in the CPU image (#61382).
- **Other**: tpu-inference v0.32.0 with torch 2.14.0 (#63080) and more out-of-tree platform hooks (#62026, #63017, #63601, #63457).

## Bug Fixes
- **DeepSeek-V4 and V4.1**: length-aware decode indexer block selection (about 850 us saved per step at 1M context, #61039); MegaMoE experts loaded directly in the DeepGEMM layout so `reload_weights` works (#60096); `fused_wqa_wkv` split by token across TP on SM90 (#58872); MegaAttn `wo_a` writes MXFP8 input for `wo_b` (#60118); native small-head FlashMLA sparse prefill on SM90 (#61615); FlashMLA interleaved KV layout (#61371); redundant EPLB expert slots loaded (#59611); GPU Engram tables read fr
- **Multimodal**: PaddleOCR-VL batched vision encoder (34-85% more throughput, #61458) and `compile_mm_encoder` (#61838); DeepSeek-VL2 ViT full CUDA graphs (#46005); Gemma4Unified batched embedding (#60643); Pixtral real-valued RoPE under compile (#63145); and fixes for EXAONE-4.5 images (broken since v0.29.0, #61104), Gemma 4 encoder graphs (#61365, #60687, #63538), DeepSeek-VL2 encoder cache and PP (#63120, #63192) and Gemma 4 placeholders (#61196).
- **Speculative decoding**: XPress (#54448), FLy (#53987), no draft-slot budget reservation for DFlash/DSpark (#59468), quantized draft `lm_head` (#59973), lookahead block hashes (#61589), MRV2 skipping drafts no request can use (up to 7.3% prefill throughput with DFlash, #62784), Triton split-K verify (#56148), varlen Mamba2 decode for adaptive verification (#53036), and fixes for stop-trimmed token counts (#58668), block verification randomness (#61161), MTP block tables (#61174), grammar masks 
- **Model Runner V2**: batch-sharded sampling for GPT-OSS (#61257) with adaptive verification (#58466), 31-120x faster long-prompt staged writes on GB200 (#62698), encoder towers skipped on later PP stages (0.58-1.27 GiB per stage, #60238), async scheduling kept for ngram (#62521), `input_ids` on later PP stages (#61016), and memory-profiling fixes (#57926, #58243, #61509, #63431).
- **Sampling**: Triton top-k/top-p skips the cutoff search for masked logits (up to about 8x, #61213); small-batch split top-p (1.22-2.70x on B300, #61996); and fixes for huge logits and boundary ties (#58231, #62046).
- **Startup and reliability**: concurrent local worker startup with TP > 1 (about 50s saved at TP8, #60028); shared Triton JIT compile pool (9-28% faster cold warmup, #61422); engine and worker watchdog (#55700); libgomp fork deadlock fixed (#62884); Fast Start with expert parallelism (#55490) and daemon fixes (#62583, #63451, #63088, #62734, #63506); and SHM broadcast reader memory growth fixed (#63165).