AI Index

Gemma 3 12B IT

Language Google Partly verified
  • In TextImage Out Text
  • Tool use
  • Open weights

Prices

1 offering · prices as listed
Cloudflare Workers AI Global · USD @cf/google/gemma-3-12b-it
Conditions/ 1M tokensInputOutput
Standard $0.345$0.556

Prices from Cloudflare Workers AI pricing (checked Nov 1, 2026)

Standard: the price a request with no conditions pays (the headline price elsewhere on the site).

Details

Reasoning
—
Knowledge cutoff
Aug 2024
Licence
gemma
Family
gemma
Model ID

Not stated on Google’s pages: max input, reasoning efforts, reasoning can be off, max thinking budget, structured output, built-in tools, aliases.

Run locally

Weights: google/gemma-3-12b-it
  • Parameters 12.2B
  • Dense
  • Native precision BF16
  • Layers 48
  • KV cache per token 360 KB
  • API context 128K (maker’s API)
My hardware GeForce RTX 4090 24 GB · 1008 GB/s · 64 GB DDR5 dual channel · 89.6 GB/s
On your hardware · GeForce RTX 4090
Runs well Q8_0 (automatic) 8K context 32–47 tok/sOne GPU
WeightsKV cacheOverheadUsable
Needs
≈ 15.5 GB (weights 12.1 + KV cache 2.8 + overhead 0.6)
Usable
21.6 GB (90% of its memory)
Generation speed
32–47 tok/s
Longest context
25K (longer doesn’t fit)
First token
about 0.39–0.59 s for a 2K prompt
PrecisionWeightsNeeds at 8KOn your hardwareLongest context
BF16 (native)maker’s file22.7 GB≈ 26.6Needs system memory 6–9 tok/s—
Q8_012.1 GB≈ 15.5Runs well 32–47 tok/s25K
FP811.4 GB≈ 14.7Runs well 33–50 tok/s27K
Q6_K9.3 GB≈ 12.6Runs well 39–58 tok/s33K
Q5_K_M8.1 GB≈ 11.4Runs well 43–65 tok/s37K
Q5_07.8 GB≈ 11.1Runs well 44–66 tok/s38K
Q4_K_M6.9 GB≈ 10.2Runs well 48–73 tok/s40K
Q4_06.4 GB≈ 9.7Runs well 51–77 tok/s42K
IQ4_XS6.0 GB≈ 9.3Runs well 53–80 tok/s43K
Q3_K_Mquality drops5.5 GB≈ 8.9Runs well 56–84 tok/s44K
IQ3_XXSquality drops4.3 GB≈ 7.7Runs well 66–98 tok/s48K
Q2_Kquality drops4.3 GB≈ 7.6Runs well 66–100 tok/s48K
IQ2_XXSquality drops2.9 GB≈ 6.2Runs well 82–123 tok/s52K

Click a row to choose a precision. “Maker’s file” sizes are the maker’s own files; others are parameters × bits per weight.

Compared with other hardware

Each at the highest precision it runs. Click one for details.

Smallest that runs it (Q4_K_M, 8K): GPU GeForce RTX 5070; unified memory Apple M5 · 16 GB.

Run commands

The maker publishes no files at this precision. Choose its native precision or one of its own builds, or quantise it yourself with llama.cpp.

Memory to fine-tune
Full≈ 183.1 GBDoesn’t fit your hardware
LoRA≈ 24.5 GBDoesn’t fit your hardware
QLoRA≈ 8.1 GBFits your hardware

Batch 1, 2,048-token sequences, gradient checkpointing; LoRA and QLoRA train rank-16 adapters on the attention projections.

Estimates; facts from the maker’s Hugging Face repositories, as of 2026-11-06. How it’s estimated

Cited scores

From third-party leaderboards, not measured by us
  • LMArena Text 1,342

    #131 of 163 listed models · CI 1,332–1,351 · 3,829 votes

  • Epoch AI · GPQA Diamond 39.5%

    #147 of 168 listed models · ±2.53

  • Epoch AI · OTIS Mock AIME 2024–2025 16.7%

    #126 of 159 listed models · ±4.24

  • EQ-Bench Creative Writing v3 1,131

    #80 of 93 listed models

  • LMArena (Arena), leaderboard-dataset on Hugging Face, licensed CC BY 4.0 · License · as of Oct 28, 2026
  • Epoch AI, 'Capabilities & benchmarking'. Published online at epoch.ai. Retrieved from 'https://epoch.ai/benchmarks'. Licensed CC BY 4.0. · License · as of Nov 4, 2026
  • EQ-Bench by Samuel J. Paech (eqbench.com; EQ-bench/EQ-bench-site on GitHub), MIT License. · License · as of Oct 28, 2026
  • changes we made

Changes

All changes →
  1. Now priced on Cloudflare Workers AI Input $0.345 · Output $0.556

Sources

Google’s official policies: data use, privacy, terms →

Some values on this page are still waiting to be checked against official pages: a filled dot is verified, a hollow one is not yet (with or without a source on file). How we verify data

Google · United States · ai.google.dev

Snapshots, the same family and other models from this provider