AI Index

Gemma 4 31B IT

Language Google Partly verified Released Apr 2, 2026
  • In TextImage Out Text
  • Reasoning
  • Reasoning can be off
  • Tool use
  • Structured output
  • Open weights

Prices

0 offerings · prices as listed

No priced offering is listed for this model yet.

Details

Knowledge cutoff
Jan 2025
Licence
Apache 2.0
Family
gemma

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

Run locally

Weights: google/gemma-4-31B-it
  • Parameters 31.3B
  • Dense
  • Native precision BF16
  • Layers 60
  • KV cache per token 960 KB
  • Config positions 256K (config.json max_position_embeddings)
  • API context 250K (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 Q4_0 (automatic) 8K context 25–38 tok/sOne GPU
WeightsKV cacheOverheadUsable
Needs
≈ 19.2 GB (weights 16.4 + KV cache 2.0 + overhead 0.8)
Usable
21.6 GB (90% of its memory)
Generation speed
25–38 tok/s
Longest context
23K (longer doesn’t fit)
First token
about 1.01–1.51 s for a 2K prompt
PrecisionWeightsNeeds at 8KOn your hardwareLongest context
BF16 (native)maker’s file58.3 GB≈ 63.2Doesn’t fit—
Q8_030.9 GB≈ 34.5Needs system memory 3–4 tok/s—
FP829.1 GB≈ 32.6Needs system memory 3–5 tok/s—
Q6_K23.9 GB≈ 27.1Needs system memory 6–9 tok/s—
Q5_K_M20.7 GB≈ 23.8Needs system memory 10–15 tok/s—
Q5_020.0 GB≈ 23.1Needs system memory 13–19 tok/s—
Q4_K_M17.7 GB≈ 20.6Runs well 24–36 tok/s14K
Q4_016.4 GB≈ 19.2Runs well 25–38 tok/s23K
IQ4_XS15.5 GB≈ 18.3Runs well 27–40 tok/s29K
Q3_K_Mquality drops14.2 GB≈ 17.0Runs well 29–43 tok/s37K
IQ3_XXSquality drops11.1 GB≈ 13.7Runs well 36–53 tok/s58K
Q2_Kquality drops10.9 GB≈ 13.5Runs well 36–54 tok/s59K
IQ2_XXSquality drops7.5 GB≈ 10.0Runs well 49–74 tok/s82K

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 4090; unified memory Apple M6 (24GB or 32GB) · 32 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≈ 470.1 GBDoesn’t fit your hardware
LoRA≈ 63.0 GBDoesn’t fit your hardware
QLoRA≈ 21.1 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
  • Epoch Capabilities Index 144.2

    #83 of 118 listed models · CI 141.9–146.7

  • Epoch AI · GPQA Diamond 75.8%

    #90 of 168 listed models · minimal effort · ±3.05

  • Epoch AI · OTIS Mock AIME 2024–2025 73.3%

    #85 of 159 listed models · minimal effort · ±6.67

  • Epoch AI · SimpleQA Verified 10.4%

    #82 of 84 listed models · ±0.97

  • 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 – 5, 2026
  • EQ-Bench by Samuel J. Paech (eqbench.com; EQ-bench/EQ-bench-site on GitHub), MIT License. · License · as of Oct 24 – 28, 2026
  • LMArena (Arena), leaderboard-dataset on Hugging Face, licensed CC BY 4.0 · License · as of Sep 13, 2026
  • changes we made

Changes

All changes →

No changes recorded yet.

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