GKRootWire
AI Google Adds 'Preferred Source' Button to Help Publishers Fight AI Traffic LossesGadgets Linkdaze Launches a Smart Calendar Aimed at Running Your Whole HouseholdSecurity Popular Rust Crate arrayref Hijacked to Spread Infostealer MalwareCloud & Sysadmin GitHub Details Cause of August 17 Outage, Outlines Reliability FixesDev Tools Show HN: 'Huzzah' Proposes a Fresh Take on AI-Assisted CodingCloud & Sysadmin The Weird Science of Cooling Data Centers With UrineAI Google Adds 'Preferred Source' Button to Help Publishers Fight AI Traffic LossesGadgets Linkdaze Launches a Smart Calendar Aimed at Running Your Whole HouseholdSecurity Popular Rust Crate arrayref Hijacked to Spread Infostealer MalwareCloud & Sysadmin GitHub Details Cause of August 17 Outage, Outlines Reliability FixesDev Tools Show HN: 'Huzzah' Proposes a Fresh Take on AI-Assisted CodingCloud & Sysadmin The Weird Science of Cooling Data Centers With Urine
AI

Qwen3 27B Hits 50 Tokens/Sec at 256K Context on a Single 24GB GPU

A blog writeup shows how careful quantization and multi-token prediction squeeze massive context windows out of a compact workstation card.

A new blog post details running Qwen3's 27B-parameter model with a full 256K token context window on an RTX PRO 4000 SFF, a small-form-factor workstation GPU with just 24GB of VRAM and 432 GB/s of memory bandwidth. Despite the modest hardware, the setup achieves roughly 50 tokens per second by leveraging multi-token prediction (MTP), a technique that lets the model generate several tokens per forward pass instead of one at a time.

The result is notable because large context windows are usually memory-hungry, and 256K tokens is enough to hold entire codebases or lengthy documents in a single prompt. The author walks through the quantization and memory-management choices needed to fit both the model weights and the KV cache for that context length into 24GB, a constraint that normally forces users toward smaller models or datacenter-grade cards.

Why it matters: This kind of optimization matters because it pushes long-context, locally-run LLMs further into reach of hobbyists and small teams without needing multi-GPU rigs or cloud API bills. MTP-style techniques are becoming a key lever for squeezing more throughput out of consumer hardware, and posts like this help the community benchmark what's actually achievable outside marketing claims.

Sources: Hacker News