# vLLM

112 Hacker News commenters recommended vLLM and 56 criticized it, November 2025 to November 2026. It was recommended in 12 of 13 months. It meets the Hacker Approved badge rule.

**Verdict:** Recommended for fast, scalable LLM serving with continuous batching, multi-GPU tensor parallelism and concurrent requests; critics cite fiddly setup, version juggling, weak AMD support and bugs.

## Why people recommend it

- Continuous batching and concurrency (7 comments): "separating the web server and GPU process, continuous batching, kv caching / chunking, and a huge model library including low precision mattered more" ([source](https://news.ycombinator.com/item?id=49205405)) "do tensor parallelism and concurrent batch requests with vllm and it becomes extremely cost competitive" ([source](https://news.ycombinator.com/item?id=46976493)) "vLLM's continuous batching helps, but there's still a physical limit with shared VRAM/compute." ([source](https://news.ycombinator.com/item?id=47644369))
- Scales well across multiple GPUs (4 comments): "do tensor parallelism and concurrent batch requests with vllm and it becomes extremely cost competitive" ([source](https://news.ycombinator.com/item?id=46976493)) "I've heard that vLLM performs much better, scaling particularly better in the multi GPU case." ([source](https://news.ycombinator.com/item?id=47796553)) "The parallelism and MTP on top of better tuned kernels gave an extremely nice boost!" ([source](https://news.ycombinator.com/item?id=48619610))
- High throughput with MTP speculative decoding (6 comments): "I have also seen a 2-3x improvement in then speeds.from improvements in vllm (mainly MTP)." ([source](https://news.ycombinator.com/item?id=48778726)) "On vllm with a 5090 I get 120-180TPS with the awq 4 bit quant + MTP speculative decoding" ([source](https://news.ycombinator.com/item?id=48026555)) "The parallelism and MTP on top of better tuned kernels gave an extremely nice boost!" ([source](https://news.ycombinator.com/item?id=48619610))
- Precise tool-call parsing (3 comments): "vllm is nicer (ray, handles queueing, doesn't have the cache invalidation bug for qwen/gemma models)" ([source](https://news.ycombinator.com/item?id=48619045)) "they've ditched the Jinja templates for an in-engine parser and I've seen no issues" ([source](https://news.ycombinator.com/item?id=48918641)) "vLLMs tool parsing is much more precise than for example llama cpp or Ollama" ([source](https://news.ycombinator.com/item?id=50246895))

## What people warn about

- Poor performance on AMD hardware (4 comments): "it's drastically slower than Ollama (tested over a few weeks, always using the official AMD vLLM nightly releases)" ([source](https://news.ycombinator.com/item?id=46614106)) "patching up vllm to run one or two models you care about on AMD hardware is a much easier proposition" ([source](https://news.ycombinator.com/item?id=49258244)) "there was an issue of abysmal performance of models on Strix Halo, that is open for half a year" ([source](https://news.ycombinator.com/item?id=49385512))
- Fiddly setup and version juggling (5 comments): "It took me quite some time to figure the magic combination of versions and commits" ([source](https://news.ycombinator.com/item?id=46680992)) "It finally stabilized for me on 2 x H200 using a commit a few days before 0.28" ([source](https://news.ycombinator.com/item?id=49493188)) "it's slow to start, it's Linux only, it doesn't like WSL much, ended up with either old or nightly container builds, I more or less have given up" ([source](https://news.ycombinator.com/item?id=48589546))
- Bugs and buggy output (4 comments): "vLLM seems to have a lot more gotchas and kludged together stuff once you get outside of anything straightforward." ([source](https://news.ycombinator.com/item?id=49493543)) "there is series of bugs in the vllm middleware that prevent GLM models from getting all of their reasoning content" ([source](https://news.ycombinator.com/item?id=49794408)) "probably one of the reasons why vllm, sglang, the llamas etc. so often produce buggy output" ([source](https://news.ycombinator.com/item?id=50293282))
- Slow startup and heavy memory use (3 comments): "It took a good 4 minutes plus to load up on the 2x 3090 rig, and served a single request 3 tokens/second slower." ([source](https://news.ycombinator.com/item?id=48587529)) "I tried FP8 in vLLM and it used 110GB and then my machine started to swap when I hit it with a query." ([source](https://news.ycombinator.com/item?id=46875833)) "vLLM will just fill whatever VRAM is available, no matter the context size, or the model size" ([source](https://news.ycombinator.com/item?id=48729048))

## Alternatives hackers recommend

- [Strata](https://hackerapproved.com/p/strata.md): 31 recommend
- [DwarfStar 4](https://hackerapproved.com/p/dwarfstar.md): 26 recommend
- [SGLang](https://hackerapproved.com/p/sglang.md): 19 recommend
- [ninfer](https://hackerapproved.com/p/ninfer.md): 16 recommend
- [exo](https://hackerapproved.com/p/exo.md): 15 recommend

## Cite as

Hacker Approved, "vLLM", https://hackerapproved.com/p/vllm/ (counts from 1,795,640 Hacker News comments, November 2025 to November 2026).