Updated August 2026 Gemma 1 through Gemma 4

Gemma Model
Comparison

Gemma against Gemma — five current sizes, four generations, and a dozen specialised variants. Put any two side by side, see what actually changed between generations, and check the licence before you build on one.

Compare two models All Gemma 4 sizes Generations Licence divide

For comparisons against Llama, Qwen and closed models, see benchmarks

Generation jump · Gemma 3 27B → Gemma 4 31B

AIME 2026 20.8→89.2

Codeforces Elo 110→2150

LiveCodeBench 29.1→80.0

MMLU Pro 67.6→85.2

Context 128K→256K

One generation, seventeen months apart

Side by side

Compare any two models

Pick two and the differences are highlighted. Everything here is from Google's technical report or official model documentation.

Model A

Model B

Current generation

The five Gemma 4 sizes

One table with everything that differs between them. Speed figures are single-stream on an H100 for the server models, and on-device for the edge ones — they aren't the same measurement, and the table says which is which.

Spec	31B	26B A4B	12B	E4B	E2B
Architecture	Dense	MoE	Unified	Edge	Edge
Active params / token	31B	~3.8B	12B	4.5B	2.3B
Context window	256K	256K	256K	128K	128K
Audio & video input	—	—	Yes	Yes	Yes
Memory · BF16	69.9 GB	57.7 GB	26.7 GB	17.9 GB	11.4 GB
Memory · Q4_0	17.5 GB	14.4 GB	6.7 GB	4.5 GB	2.9 GB
Runs on a phone	—	—	—	Yes	Yes
MMLU Pro	85.2	82.6	77.2	69.4	60.0
AIME 2026	89.2	88.3	77.5	42.5	37.5
GPQA Diamond	84.3	82.3	78.8	58.6	43.4
LiveCodeBench v6	80.0	77.1	72.0	52.0	44.0
MMMU Pro vision	76.9	73.8	69.1	52.6	44.2
RULER @128K	96.4	89.8	91.2	86.6	70.4
Speed H100 / device	40 t/s	177 t/s	—	22 t/s	52 t/s
With MTP	125 t/s	264 t/s	—	—	—

🔍

Two rows worth staring at. The 26B A4B scores within about three points of the 31B on every accuracy benchmark while running 4.4× faster — that's the strongest argument in the table, and it's why the MoE is the right default for serving. And the audio row runs the opposite way to everything else: the two largest models can't take audio at all.

Over time

Four generations of Gemma

Comparing flagship to flagship. The interesting part is that the jump from 3 to 4 is far larger than any before it.

Generation	Released	Sizes	Context	Modalities	Licence
Gemma 4	Mar 31, 2026	E2B, E4B, 12B, 26B A4B, 31B	256K	Text, image, audio, video	Apache 2.0
Gemma 3	Mar 10, 2025	270M, 1B, 4B, 12B, 27B	128K	Text, image	Gemma Terms
Gemma 3n	Jun 26, 2025	E2B, E4B	32K	Text, image, audio	Gemma Terms
Gemma 2	Jun 27, 2024	2B, 9B, 27B	8K	Text	Gemma Terms
Gemma 1	Feb 21, 2024	2B, 7B	8K	Text	Gemma Terms

Gemma 3 27B against Gemma 4 31B

Benchmark	Gemma 3 27B	Gemma 4 31B	Change
AIME 2026	20.8	89.2	+68.4
Codeforces Elo	110	2150	+2040
LiveCodeBench v6	29.1	80.0	+50.9
BBH	19.3	74.4	+55.1
GPQA Diamond	42.4	84.3	+41.9
MMLU Pro	67.6	85.2	+17.6
MMMU Pro	49.7	76.9	+27.2
MATH-Vision	46.0	85.6	+39.6
RULER @128K	66.0	96.4	+30.4
LOFT retrieval	8.6	79.5	+70.9
IFEval	90.4	98.9	+8.5
LMArena Elo	1366	1451	+85

Where the jump came from

Almost all of it is reasoning-focused post-training. AIME going from 20.8 to 89.2 and Codeforces from 110 to 2150 isn't a scaling result — the models are a similar size. It's thinking mode plus training that teaches the model to use it. The long-context numbers tell a similar story: LOFT retrieval went from a near-useless 8.6 to 79.5.

The small models moved most

Gemma 4 E2B — two billion parameters, running on a phone — scores 37.5 on AIME against Gemma 3 27B's 20.8, and 44.0 on LiveCodeBench against 29.1. A model you can fit in about a gigabyte now outperforms the previous generation's flagship on reasoning and code.

Read this before shipping

The licence divide

This is the single biggest practical difference between Gemma 4 and everything before it, and it doesn't show up in any benchmark.

Gemma 4 only

Apache 2.0

✅ Commercial use, no conditions attached
✅ Modify, fine-tune and redistribute freely
✅ Build proprietary products on top
✅ No use restrictions to pass downstream
✅ Standard, well-understood terms your legal team already knows
✅ Google cannot unilaterally change the terms later

Applies to the Gemma 4 family. Google publishes it separately from the older Gemma Terms.

Gemma 1, 2, 3, 3n and most variants

Gemma Terms of Use

⚠️ A Prohibited Use Policy applies, incorporated by reference
⚠️ You must pass the restrictions on to anyone you redistribute to
⚠️ You must supply recipients with a copy of the full agreement
⚠️ Google reserves the right to restrict usage remotely
⚠️ No use of Google trademarks or implied endorsement
⚠️ Terms can be updated by Google over time

Covers Gemma 1, 1.1, 2, 3 and 3n, plus CodeGemma, PaliGemma, ShieldGemma, EmbeddingGemma, FunctionGemma and others.

⚖️

The obligation propagates, and that's what catches people. Under the Gemma Terms, if you fine-tune Gemma 3 and ship the result, you have to carry the use restrictions forward to your users and give them the agreement. Apache 2.0 imposes no such duty. If you're choosing between a Gemma 3 fine-tune you already have and moving to Gemma 4, the licence alone is often the deciding factor — before you even look at the benchmark gap. This is a summary, not legal advice: read the actual terms for your situation.

Practical

Should you move off Gemma 3?

✅

Yes, clearly

If you're doing anything involving reasoning, maths, code or long context, the gap is enormous — and the Apache 2.0 licence removes downstream obligations you currently carry. For most projects this isn't a close call.

🤔

Weigh it up

If you have a heavily fine-tuned Gemma 3 that works, you'd be redoing that work. Budget for it — but note that a Gemma 4 model two sizes smaller may beat your tuned Gemma 3, which can cut serving costs enough to pay for the migration.

⏸️

Maybe not yet

Some specialised variants — PaliGemma, ShieldGemma, CodeGemma — have no Gemma 4 equivalent. If you depend on one, you stay on the older terms until it's updated.

📉

Consider dropping a size while you're at it. Gemma 4 12B beats Gemma 3 27B comfortably across the board — 77.2 against 67.6 on MMLU Pro, 77.5 against 20.8 on AIME — at roughly a quarter of the memory. Migrating from Gemma 3 27B to Gemma 4 12B is an upgrade in quality and a large reduction in serving cost simultaneously.

Specialised

Variants, and which generation they're on

The wider family is large and unevenly updated. Most of these are still on the older licence.

Variant	Released	Sizes	Purpose	Licence
DiffusionGemma	Jun 2026	26B A4B	Text diffusion — parallel generation, ~4× faster output	Apache 2.0
TranslateGemma	Jan 15, 2026	4B, 12B, 27B	Dedicated machine translation	Gemma Terms
MedGemma 1.5	Jan 13, 2026	4B	Medical imaging and clinical text	Gemma Terms
FunctionGemma	Dec 18, 2025	270M	Tiny tool-calling router	Gemma Terms
T5Gemma v2	Dec 18, 2025	270M, 1B, 4B	Encoder-decoder tasks	Gemma Terms
VaultGemma	Sep 13, 2025	1B	Differentially private training	Gemma Terms
EmbeddingGemma	Sep 4, 2025	308M	Retrieval and RAG embeddings	Gemma Terms
PaliGemma 2	Feb 19, 2025	3B, 10B, 28B	Vision-language	Gemma Terms
ShieldGemma 2	Jul 31, 2024	4B	Content safety classification	Gemma Terms
CodeGemma	Apr 9, 2024	2B, 7B	Code completion — superseded by Gemma 4	Gemma Terms
RecurrentGemma	Apr 9, 2024	2B, 9B	Griffin architecture, not a transformer	Gemma Terms

💡

CodeGemma is not the coding model any more. It dates from April 2024 and Gemma 4 comfortably surpasses it — the 26B A4B scores 77.1 on LiveCodeBench v6 against Gemma 3 27B's 29.1, and CodeGemma is two generations older still. Use plain Gemma 4 for code. The variants worth reaching for are the ones doing something Gemma 4 genuinely doesn't: translation, medical, embeddings, safety classification.

FAQ

Common questions

Which Gemma 4 model should I pick?

26B A4B for serving — near-flagship quality at 4.4× the speed. 31B when quality decides the outcome. 12B for a single consumer GPU, or for anything involving audio or video, since the larger models don't accept them. E4B and E2B for phones and edge devices.

Is the 31B actually worth it over the 26B?

Only sometimes. The accuracy gap is roughly two to three points on most benchmarks while the 31B runs at about a quarter of the speed at concurrency 1. If you're serving interactive users, the MoE is almost certainly the better trade. Reserve the 31B for offline batch work or cases where the last few points genuinely matter.

Are Gemma 3 models still supported?

They still work and remain downloadable, but they're on the older Gemma Terms of Use with its downstream obligations, and the capability gap is large. New projects should start on Gemma 4 unless they depend on a specialised variant that hasn't been updated.

What's the difference between Gemma 3n and Gemma 4 E2B/E4B?

Gemma 3n was the previous on-device family, and Gemma 4's E-series replaces it. Gemma 4 improves speech translation by 12% for E2B and 10% for E4B, and transcription by 17% and 12% — while shrinking the audio encoder from 680M to 305M parameters. Gemma 4 also brings the Apache 2.0 licence, which 3n doesn't have.

Can I still use CodeGemma or PaliGemma?

Yes, but think carefully. Both predate Gemma 4 by two generations, and plain Gemma 4 outperforms CodeGemma at coding and handles vision better than PaliGemma for most tasks. They also carry the older licence. The variants still worth using are those covering ground Gemma 4 doesn't — TranslateGemma, MedGemma, EmbeddingGemma, ShieldGemma.

Why does Gemma 4 not have a 1B or 270M model?

The E-series fills that role differently. Rather than a small dense model, E2B uses mixed 2/4/8-bit weights so its text-only footprint can drop to roughly 0.8 GB — comparable in memory to a tiny model, but far more capable. If you need something genuinely tiny for a narrow job, FunctionGemma at 270M or EmbeddingGemma at 308M are the specialised options.

Does the licence change apply retroactively to Gemma 3?

No. Google's Gemma Terms of Use still lists Gemma 1, 1.1, 2, 3 and 3n along with most specialised variants, and points to a separate Apache 2.0 licence for Gemma 4 only. Older models stay on the older terms.

Picked one?

Exact repository IDs and pull commands for every checkpoint in the family.

Go to downloads Full benchmark tables Capabilities detail