What hardware do you need to run Gemma 4 12B?

Gemma 4 12B at Q4_K_M is 7.1 GB of weights, with a context ceiling of 256k tokens. Not yet rated: released after our last ratings pass. Google's mid-size Gemma 4. The windowed layers keep the context nearly free.

Cheapest machine that runs itMac mini M6, 16GB at $899

Fastest of the ones listedMac Studio M5 Ultra, 96GB — 113 tok/s at 32k context

Honest answer on costPays back in 71 years at 500k tokens a day.

How good is it, really?

On the Artificial Analysis Intelligence Index v4.3 it scores 14 (reasoning mode; 9 without), which puts it in the Below every hosted tier band. Fine for simple, well-specified tasks. Noticeably less capable than anything the big labs sell today. Score source. See the whole table.

Summarising — not rated
Translation — not rated
Everyday coding — not rated
Reasoning & maths — not rated
Agentic work — not rated

What it costs either way

Nobody rents Gemma 4 12B by the token. The closest hosted match, Qwen3.5 9B, costs $0.08 per million input tokens and $0.13 per million output (OpenRouter, cheapest active endpoint, checked 2026-09-15). Buying a machine only beats that if you use it hard enough, for long enough, that the hardware price divides down below the rental bill.

Machines that run it

Machine	Price	Speed at 32k	Pay-back
Mac mini M6, 16GB	$899	14 tok/s estimated	Pays back in 71 years	Run the numbers
Strix Halo Framework Desktop, 32GB	$1,269	24 tok/s estimated	Pays back in 104 years	Run the numbers
MacBook Air M5 (13-inch), 16GB	$1,299	14 tok/s estimated	Pays back in 102 years	Run the numbers
MacBook Pro M5 (14-inch), 16GB	$1,999	14 tok/s estimated	Pays back in 157 years	Run the numbers
Mac Studio M5 Max, 36GB	$2,499	43 tok/s estimated	Pays back in 187 years	Run the numbers
DGX Spark GB10 Grace Blackwell, 128GB	$4,699	26 tok/s estimated	Pays back in 391 years	Run the numbers

One machine per family, cheapest first. Speeds are measured where a public benchmark exists and estimated from memory bandwidth otherwise; the calculator says which for any configuration.

The specifics

Parameters: 12B
Quantisation: Q4_K_M
Weights on disk: 7.1 GB
KV cache: 0.9 GB at 32k context — 8 of 48 layers grow with the context, and those use a different geometry from the other 40: one KV head at 512 wide (2,048 B/token) against eight heads at 256 wide on the windowed layers, which stop growing at 1,024 tokens.
Maximum context: 256k tokens (256k)
Licence: Apache 2.0
Sources: source 1, source 2