---
title: "donald991234 (@donald991234) | Hashnode"
description: "donald991234 (@donald991234) writes on Hashnode. Read their latest articles and follow their work."
url: "https://hashnode.com/@donald991234"
---

@donald991234

# donald991234

@donald991234 · Joined September 2026

### [Qwen3.8 27B VRAM, measured: Q4\_K\_M uses 19.95 GiB at 32K context](https://modelvram.hashnode.dev/qwen3-8-27b-vram-measured-q4-k-m-uses-19-95-gib-at-32k-context?utm_source=hashnode&utm_medium=feed)

Real llama.cpp numbers, why the KV cache is only 64 KiB per token, and which GGUF fits a 16, 24 or 32 GB GPU

5 min read · [modelvram.hashnode.dev](https://modelvram.hashnode.dev/qwen3-8-27b-vram-measured-q4-k-m-uses-19-95-gib-at-32k-context?utm_source=hashnode&utm_medium=feed)

[Oct 12](https://hashnode.com/posts/qwen3-8-27b-vram-measured-q4-k-m-uses-19-95-gib-at-32k-context/6accd6017fa5210e386410e5)

### [How much extra VRAM speculative decoding needs in llama.cpp (MTP, DFlash, draft models)](https://modelvram.hashnode.dev/how-much-extra-vram-speculative-decoding-needs-in-llama-cpp-mtp-dflash-draft-models?utm_source=hashnode&utm_medium=feed)

Speculative decoding makes local LLMs noticeably faster, but it is not free: the draft (an MTP head, a DFlash drafter or a small draft model) needs its own VRAM. People regularly run out of memory aft

4 min read · [modelvram.hashnode.dev](https://modelvram.hashnode.dev/how-much-extra-vram-speculative-decoding-needs-in-llama-cpp-mtp-dflash-draft-models?utm_source=hashnode&utm_medium=feed)

[4](https://hashnode.com/posts/how-much-extra-vram-speculative-decoding-needs-in-llama-cpp-mtp-dflash-draft-models/6abb4e53f601645d75bc6149) [Sep 29](https://hashnode.com/posts/how-much-extra-vram-speculative-decoding-needs-in-llama-cpp-mtp-dflash-draft-models/6abb4e53f601645d75bc6149)
