Only with CPU offload

Gemma 4 31B at Q4_K_M needs 20.2 GB but only 7.0 GB is addressable, so about 75% of the layers would stream from system RAM at roughly 60 GB/s. Expect around 2.6 tokens per second — usable for batch work, painful for chat.

25% on GPU 8K context Q4_K_M · 17.6 GB Apache 2.0 Released 2 Apr 2026 Vision

The dense flagship: strongest maths of the 24–32 GB class (89% AIME), clean prose, vision. Q4 is a tight 24 GB fit.

Hardware requirements What hardware do I need to run Gemma 4 31B? →

At Q4_K_M with an 8K context: a 32 GB GPU (RTX 5090) or a 32 GB Mac (M4). Minimum, recommended, high-quality and long-context tiers, with all 114 devices checked at every quantisation.

Fits instead: Gemma 4 E4B (5.2 GB) · Gemma 4 E2B (3.5 GB)

The VRAM budget

weights 17.6 GB
Weights 17.6 GB KV cache @ 8K 2.03 GB Runtime overhead 0.6 GB Over budget 13.2 GB past 7.0 GB

Quantisation ladder

QuantWeightsTotal @ 8KMax contextTok/sQualityFit
Q8_0 31.0 GB 33.6 GB — ~1.3 −0.1% ppl 26.6 GB over
Q6_K 23.9 GB 26.5 GB — ~1.8 −0.4% ppl 19.5 GB over
Q5_K_M 20.7 GB 23.3 GB — ~2.2 −0.8% ppl 16.3 GB over
Q4_K_M 17.6 GB 20.2 GB — ~2.6 −1.9% ppl 13.2 GB over
Q3_K_M 14.2 GB 16.9 GB — ~3.5 −5.4% ppl 9.9 GB over
Q2_K 12.2 GB 14.8 GB — ~4.3 −15% ppl 7.8 GB over

Quality is the published perplexity delta against f16 weights. Max context assumes an f16 KV cache; q8_0 roughly doubles it. This model interleaves sliding-window layers (1024 tokens, 10 of 60 layers global), which is why its cache barely grows with context.

How to run it

terminal
$ llama-server \
    -hf google/gemma-4-31B-it:Q4_K_M \
    -c 8192 -ngl 14

The engine underneath most of the others. Every knob is exposed. More on llama.cpp.

01Download is 17.6 GB. Keep it on an SSD — a first load off a spinning disk takes minutes.
02Close anything else holding VRAM. A browser with hardware acceleration can sit on 1–2 GB.
03If it falls back to CPU silently, drop the context first, then step down a quantisation.
See all models for this rig Compare with another model

Related models on an Arc A750

ModelTotalTok/s est.
Gemma 4 26B-A4B
26.5B MoE · same family
Q4_K_M · 16.0 GB ~10 CPU offload
Gemma 4 12B
12B · same family
Q4_K_M · 8.2 GB ~20 CPU offload
Gemma 4 E4B
8.0B · same family
Q4_K_M · 5.2 GB 69 Runs great
Gemma 4 E2B
5.1B · same family
Q4_K_M · 3.5 GB 108 Runs great
GLM-4.7-Flash 30B-A3B
31.2B MoE · similar size
Q4_K_M · 18.6 GB ~12 CPU offload
Nemotron 3.5 Lightning 30B-A3B
31.6B MoE · similar size
Q4_K_M · 18.4 GB ~11 CPU offload
Qwen3 30B-A3B
30.5B MoE · similar size
Q4_K_M · 18.5 GB ~11 CPU offload

Every model against the Arc A750 →

Gemma 4 31B on other hardware

DeviceTotalTok/s est.
GeForce RTX 4060 Laptop
7.0 GB usable · 256 GB/s
20.2 GB ~2.5 CPU offload
GeForce RTX 3060 12 GB
10.6 GB usable · 360 GB/s
20.2 GB ~3.3 CPU offload
M4 · 16 GB
10.7 GB usable · 120 GB/s
20.2 GB — Won't fit
GeForce RTX 4090
22.4 GB usable · 1008 GB/s
20.2 GB 35 Tight fit
Radeon RX 7900 XTX
22.4 GB usable · 960 GB/s
20.2 GB 33 Tight fit
GeForce RTX 5090
30.4 GB usable · 1792 GB/s
20.2 GB 62 Runs great
M4 · 32 GB
24.0 GB usable · 120 GB/s
20.2 GB 3.8 Runs great
M3 Max · 64 GB
48.0 GB usable · 400 GB/s
20.2 GB 13 Runs great

All 114 devices, by tier →