Historical brief · 2026-11-17

vLLM v0.34.0: Major Performance Boosts and New Model Support

vLLM v0.34.0 brings massive speedups for running large language models, alongside native support for 10 new models like Mistral Large 4 and Nemotron-3.5-Transcribe. This update turns on the extensible KV cache by default to give your GPUs a bigger memory cache and higher overall throughput.

Source image · vLLM Releases

Release Overview

vLLM v0.34.0 marks a significant milestone driven by 1,061 commits from 412 contributors. This release completely phases out Model Runner V1, making Model Runner V2 the exclusive GPU runner for maximum efficiency. It also introduces major speedups tailored for Blackwell and MI355X hardware.

Core Performance Improvements

Performance is a major focus in this version. Users can now experience up to 1.95× decode acceleration on GB200 systems, alongside a 1.6× throughput boost for models like GLM-5.3-Flash. Additionally, long-prompt Time-to-First-Token (TTFT) has been slashed, running 1.24 to 1.40 times faster.

New Model Integrations

The update adds seamless support for 10 new state-of-the-art models, covering text, speech, and vision modalities. Highlights include Mistral Large 4 (featuring FP8 and NVFP4 support), MiniMax-M3.1-Flash, NVIDIA Nemotron-3.5-Transcribe, and Liquid AI's LFM3 series.

Breaking Changes & Upgrades

Several important changes require attention before upgrading. The default block size is now 64 instead of 16, and huggingface_hub >= 2.0.0 is strictly required. Furthermore, the extensible KV cache now automatically sizes itself based on actually free memory after warmup, maximizing your hardware's potential.

View selection rationale
Open sourceStage not established

Importance 88/100

Signal90Structure90Consequence85Confidence93Momentum92
About this brief
Content type
News summary
Source type
First-party / official source
Original source
vLLM Releases ↗
Source published
Process
AI-assisted discovery, translation, summarisation, classification and scoring; standards and corrections are maintained by the editorial desk.
Verification
The original URL and publication metadata are retained. This brief is not independently fact-checked unless explicitly stated.
Published by
Polyemergence Editorial Desk · PolyEmergence Pte Ltd.
Conflict disclosure
No financial or commercial conflict declared.

Editorial standards → · Corrections →

POLYEMERGENCE / EXPLORE

Keep exploring

Follow the topic, not just the headline.

Explore the Artificial Intelligence archive →