Models & Labs

vLLM v0.30.0: DeepSeek V4.1 and Fast Start

vLLM ReleasesSeptember 22, 2026high confidence

Why it matters

→DeepSeek-V4.1-Flash support allows MXFP8 KV caching on SM100, significantly reducing memory footprint for large reasoning models.
→Fast Start daemon enables near-instant engine restarts by caching weights in GPU memory via CUDA IPC, eliminating disk I/O bottlenecks.
→HiSparse host-resident tier spills KV pages to pinned host memory, extending effective context windows without additional GPU hardware.

vLLM has released version 0.30.0, featuring support for DeepSeek-V4.1-Flash and GLM-5.3-Flash. Key performance improvements include Fast Start, a persistent weight-cache daemon that reduces engine initialization time by mapping weights via CUDA IPC, and HiSparse, which offloads sparse MLA KV pages to host memory under GPU pressure. The release also introduces Gumbel-max watermarking for AI detection and updates the Model Runner V2 with dual-batch overlap in eager mode. Breaking changes include making scale-out endpoints opt-in and removing deprecated environment variables.

Read original

The story around this

Earlier coverage that leads up to this article, and what followed. Lines connect each piece to the closest one after it, converging here.

How we got here

1
vLLM Boosts Transformers Backend for Native-Speed Inference
Hugging Face Blog · July 8, 2026 · Background
2
llama.cpp b9952 Release Enhances DeepSeek V4
llama.cpp Releases · July 11, 2026 · Related
3
DeepSeek V4 Flash Released
Lev Selector · July 31, 2026 · Related
4
DeepSeek-V4 Flash vs GPT-5.6 Luna: Cost vs Quality
Together AI Blog · August 6, 2026 · Background
5
DeepSeek V4 Pro 0813 vs Claude Fable 5: Cost and Efficiency
Together AI Blog · August 17, 2026 · Related
6
vLLM v0.28.0 Release: Major Performance Enhancements
vLLM Releases · August 28, 2026 · Same story
7
DeepSeek Releases DeepSeek-V4.1-Flash Model
Matt Wolfe · September 11, 2026 · Related
8
DeepSeek Releases V4.1 Flash Model
Lev Selector · September 11, 2026 · Related

What happened next

9
Liquid AI releases DSpark drafter for vision models
Hugging Face Blog · September 24, 2026 · Related
10
vLLM v0.31.0: SM100 defaults and fast restarts
vLLM Releases · October 6, 2026 · Same story
11
llama.cpp deepseekv4.1 follow-up fixes
llama.cpp Releases · October 26, 2026 · Same story

More from vLLM Releases

Models & Labsmodels

vLLM v0.34.0rc0 adds Gemma and StableLM support

One of llama.cpp's quietest but biggest releases. With ROCm 7.2 added as a default build, AMD GPU users stop being second-class citizens for local inference — the gap with CUDA narrows visibly. Apple Silicon Macs now compile in KleidiAI by default, meaning every M-series machine gets ARM-tuned GEMM kernels for free, no flag-flipping required. There's no new model and no new quantization here — just llama.cpp quietly becoming the inference runtime for everyone who isn't on NVIDIA. That's the headline.

vLLM ReleasesNov 11, 2026

More in Models & Labs

Models & Labsother

llama.cpp b12340 adds CUDA 13 and ROCm 10.0

This release quietly closes the hardware gap for local inference by adding default support for CUDA 13 and ROCm 10.0 across Linux and Windows. Users with newer NVIDIA GPUs or AMD accelerators no longer need to compile from source to get performance, while Apple Silicon builds now ship KleidiAI disabled by default, shifting the focus back to standard ARM optimizations. The inclusion of Snapdragon support for Linux and Android marks a significant step toward on-device AI beyond the x86/ARM desktop ecosystem. It is less about new features and more about ensuring llama.cpp remains the universal runtime for every available accelerator.

llama.cpp ReleasesNov 13, 2026

Models & Labscoding

llama.cpp b12342 adds few-row MMQ for speculative decoding

Speculative decoding just got a significant speed boost on NVIDIA GPUs. This release introduces a specialized kernel for 'few-row' matrix multiplications, targeting the small batch sizes typical of verification passes in speculative decoding workflows. By optimizing these specific operations for quantized formats like Q8_0 and IQ4 variants, llama.cpp reduces latency where it matters most during token generation. The change is CUDA-only for now but leverages hardware capabilities from Ampere onwards, making local inference faster without requiring new model architectures or complex configuration changes.

llama.cpp ReleasesNov 13, 2026

Models & Labsmodels

Google deploys Gemini 4 Hydrogen for faster Search

Google is quietly shifting its search infrastructure by replacing older fast models with Gemini 4 Hydrogen, a tier designed specifically for speed and cost-efficiency rather than raw reasoning power. This model completes tasks 2.5 times faster than its predecessor while beating the more powerful Argon on specific intelligence benchmarks, proving that specialized efficiency can outperform brute force in high-volume workflows. By integrating this into Search's fan-out technique, Google now breaks down complex queries into subqueries with greater precision and speed, delivering richer visual responses without the latency of heavier models. This marks a strategic pivot toward agentic execution layers within consumer search, making AI Mode feel snappier for everyday multi-part questions.

TechCrunch AINov 12, 2026