vLLM v0.30.0 ships DeepSeek-V4.1-Flash support, weight-cache fast start, and watermarking in 762-commit release

vLLM v0.27.0 ships with 561 commits from 242 contributors, adding full-stack support for Kimi K3 including kernels, frontends, and quantization. It also introduces Qwen3.5 models and K-EXAONE-2.0-750B, expanding the inference engine's model coverage.

Social·Community

DeepSeek-V4.1-Flash (552B MoE, 1M ctx) served on four DGX Sparks via vLLM TP4 at 73.8 tok/s decode

A GitHub project documents running DeepSeek-V4.1-Flash (552B MoE, MXFP4 experts, 1M context) on four NVIDIA DGX Sparks using vLLM with TP4, including an Engram-on-disk patch and sm121 kernel builds. The repo reports measured decode throughput, concurrency results, prefill speeds, and passing end-to-end vision and tool-calling checks. It offers practitioners a concrete low-end multi-node serving recipe within hours of the model dropping.

A GitHub project documents running DeepSeek-V4.1-Flash (552B MoE, MXFP4 experts, 1M context) on four NVIDIA DGX Sparks using vLLM with TP4, including an Engram-on-disk patch and sm121 kernel builds. The repo reports measured decode throughput, concurrency results, prefill speeds, and passing end-to-end vision and tool-calling checks. It offers practitioners a concrete low-end multi-node serving recipe within hours of the model dropping.

Releases·Model Releasefeatured

DeepSeek V4.1 Flash is getting surprisingly close to GPT-5.6 Sol territory, while being absurdly cheap

DeepSeek released V4.1-Flash, an open-weight (MIT) MoE model built on a new causal encoder-decoder architecture with asymmetric active parameters — 8B for input/prefill and 16B for output/decode. The release adds vision input, 1M-token context, and aggressive KV-cache compression, drawing independent praise as the new top open-weight model on cost-adjusted intelligence. It succeeds V4 Pro, which DeepSeek is quietly retiring, and several commentators argue the '0.1 version' label understates a fundamental architectural overhaul.

DeepSeek released V4.1-Flash, an open-weight (MIT) MoE model built on a new causal encoder-decoder architecture with asymmetric active parameters — 8B for input/prefill and 16B for output/decode. The release adds vision input, 1M-token context, and aggressive KV-cache compression, drawing independent praise as the new top open-weight model on cost-adjusted intelligence. It succeeds V4 Pro, which DeepSeek is quietly retiring, and several commentators argue the '0.1 version' label understates a fundamental architectural overhaul.

Social·Community

Users run GLM-5.3-Flash at 1M+ context on DGX Spark, GB300, and RTX6000 Pro rigs

Z.ai released GLM-5.3-Flash (320B/18B-active) on 2026-08-26, and community members immediately published vLLM serving recipes for it on local hardware. A GitHub repo demonstrates 1,048,576-token context on two DGX Spark nodes using NVFP4 weights and KV cache, while Reddit users report 206 tok/s on a DGX Station GB300 and 1.4M context on 4x RTX6000 Pro.

Z.ai released GLM-5.3-Flash (320B/18B-active) on 2026-08-26, and community members immediately published vLLM serving recipes for it on local hardware. A GitHub repo demonstrates 1,048,576-token context on two DGX Spark nodes using NVFP4 weights and KV cache, while Reddit users report 206 tok/s on a DGX Station GB300 and 1.4M context on 4x RTX6000 Pro.

Releases·Tool Release

Serving GLM-5.3-Flash on 8×H100 in 37s with vLLM snapshots

A developer released warmctl, an open-source CLI that snapshots a warmed-up vLLM engine and restores it in seconds on a different host with multi-GPU support, demoing GLM-5.3-Flash on 8×H100 starting in 37 seconds. It enables scale-to-zero deployments without the ~10-minute cold start. Similar capabilities exist in hosted services (Modal, Inferize, RunPod FlashBoot), but warmctl runs on users' own GPUs.

A developer released warmctl, an open-source CLI that snapshots a warmed-up vLLM engine and restores it in seconds on a different host with multi-GPU support, demoing GLM-5.3-Flash on 8×H100 starting in 37 seconds. It enables scale-to-zero deployments without the ~10-minute cold start. Similar capabilities exist in hosted services (Modal, Inferize, RunPod FlashBoot), but warmctl runs on users' own GPUs.