Google's open-weight family — strong instruction-following at small parameter counts. Available variants: 2B 7B 9B 27B.
As an Amazon Associate, SpecPicks earns from qualifying purchases. See our review methodology.
Gemma 4 punches well above its weight class on instruction-following and conversational quality benchmarks. The 9B variant is the best "fits in 12 GB VRAM and feels like a much bigger model" pick in the 2023 open-weight space — perfect for a daily-driver chat rig on consumer hardware.
Gemma 4 2B at Q4 fits even on integrated graphics. 7B Q4 needs 8 GB; 12 GB lets you keep full 128K context active.
*Price sourced from Amazon.com. Price and availability subject to change.
9B is the daily-driver sweet spot. The 27B variant fits at Q4 on 16 GB but context room gets tight; 24 GB (RTX 4090) is the smooth-running tier.
*Price sourced from Amazon.com. Price and availability subject to change.
Gemma 4 27B at Q8 is the highest-quality run that still fits a single consumer card. RTX 5090 (32 GB) lets you run Q8 + full 128K context comfortably.
*Price sourced from Amazon.com. Price and availability subject to change.
For chat / writing quality at small sizes, Gemma 4 9B beats Llama 4's 8B variant on most leaderboards. For coding or tool-use, Llama 4 / Qwen3 are stronger. For pure conversational quality on a 12-16 GB card, Gemma 4 9B is the choice.
The 9B and 27B variants have native vision support in v4.1. The 2B and 7B base variants are text-only.
Used RTX 3090 ($800-1100). 24 GB VRAM holds 27B at Q4 with comfortable context. RTX 4090 / 5090 give better tok/s but the 3090 is the price/perf winner.