Canonical datasheet Updated August 2026

Gemma 4 Specs

Every published number for the Gemma 4 family in one place - architecture, memory, benchmarks, throughput, on-device performance and provenance. This is the reference the other pages link to rather than restate.

Core specifications Benchmarks Throughput Provenance

Every figure is attributed in provenance - self-reported and independent are marked separately

Family at a glance

Models5 sizes

Parameters2.3B – 31B

Context128K / 256K

Vocabulary262,144

Languages140+

Data cutoffJanuary 2025

LicenceApache 2.0

Released March 31, 2026 · 12B added June 3 · refreshed July 15

📋 Core 💾 Memory 🧱 Architecture 📊 Text benchmarks 👁️ Vision 📜 Long context 🎙️ Audio ⚡ Throughput 📱 On-device 📦 Checkpoints 🔍 Provenance

Core

Model specifications

Specification	31B	26B A4B	12B	E4B	E2B
Total parameters	31B	26B	11.95B	4.5B	2.3B
Active per token	31B	~3.8B	11.95B	4.5B	2.3B
Architecture	Dense	MoE	Unified	Dense edge	Dense edge
Context window	256K	256K	256K	128K	128K
Text input	✓	✓	✓	✓	✓
Image input	✓	✓	✓	✓	✓
Audio input	-	-	✓	✓	✓
Video input	-	-	✓	✓	✓
Output	Text only - no model in the family generates images or audio
Thinking mode	✓	✓	✓	✓	✓
Function calling	✓	✓	✓	✓	✓
Runs on a phone	-	-	-	✓	✓
MTP drafter published	✓	✓	✓	✓	✓

Memory

Weights by precision

Weights only. Add roughly 20–30% for KV cache at moderate context; more as context and concurrency grow.

Model	BF16	SFP8	Q4_0	Mobile	Context
Gemma 4 31B	69.9 GB	34.9 GB	17.5 GB	-	256K
Gemma 4 26B A4B	57.7 GB	28.8 GB	14.4 GB	-	256K
Gemma 4 12B	26.7 GB	13.4 GB	6.7 GB	-	256K
Gemma 4 E4B	17.9 GB	8.9 GB	4.5 GB	2.5 GB	128K
Gemma 4 E2B	11.4 GB	5.7 GB	2.9 GB	1.1 GB	128K

E2B's text-only footprint drops to roughly 0.8 GB with mixed 2/4/8-bit weights. Community Q4_K_M quants run about 10% larger than the official Q4_0 figures.

Architecture

Implementation detail

Explained in context on the technical page - this is the reference table.

Specification	31B	26B A4B	12B	E4B	E2B
Base architecture	Decoder-only Transformer
Normalisation	RMSNorm, pre + post · QKNorm on queries and keys
Tokenizer	SentencePiece, 262,144 entries · split digits · preserved whitespace · byte fallback
Local:global attention	5:1	5:1	5:1	5:1	4:1
Global positional	pp-RoPE, p = 0.25, base frequency 1M
Local positional	Standard RoPE, base frequency 10k
Global KV reduction	37.5% - keys reused as values in global attention layers
KV cache sharing	-	-	-	18/42	20/35
Vision path	550M ViT	550M ViT	35M projection	150M ViT p16	150M ViT p16
Vision patch size	-	-	48×48×3 RGB	16	16
Vision token budget	70 · 140 · 280 (default) · 560 · 1120
Audio path	-	-	none - raw projection	305M USM	305M USM
Audio frame	-	-	40 ms @ 16 kHz → 640-dim	40 ms Mel	40 ms Mel
MTP drafter layers	4	4	4	4	4
MTP drafter dim	1024	1024	~0.4B params	256	256
MTP vocab projection	Clustered from 262k to 4,096 - roughly 64× reduction
Training hardware	TPUv5p and TPUv6e · ZeRO-3 sharding · Pathways · JAX / GSPMD / MegaScale XLA
Data cutoff	January 2025

Text

Reasoning and code

Instruction-tuned checkpoints with thinking mode enabled. Gemma 3 27B included as the generational baseline.

Benchmark	31B	26B A4B	12B	E4B	E2B	Gemma 3 27B
MMLU Pro	85.2	82.6	77.2	69.4	60.0	67.6
AIME 2026 no tools	89.2	88.3	77.5	42.5	37.5	20.8
GPQA Diamond	84.3	82.3	78.8	58.6	43.4	42.4
LiveCodeBench v6	80.0	77.1	72.0	52.0	44.0	29.1
Codeforces Elo	2150	1718	1659	940	633	110
BBH micro avg	74.4	64.8	53.0	33.1	21.9	19.3
SciCode	43.0	40.0	38.0	24.0	21.0	21.0
IFEval	98.9	98.5	97.2	96.7	94.6	90.4
IFBench	76.0	72.0	74.0	44.0	38.0	32.0
HLE	19.5	8.7	5.2	-	-	-
HLE with search	26.5	17.2	-	-	-	-

Human preference · LMArena text

Model	Elo	Rank	Category	As of
Gemma 4 31B	1451 ± 8	43	Leading open dense model	Jun 19, 2026
Gemma 4 26B A4B	1438 ± 8	61	Open MoE	Jun 19, 2026
Gemma 3 27B	1366 ± 4	157	Previous generation	Jun 19, 2026

At launch LMArena placed the 31B at #3 among open models and #27 overall. Arena ranks decay as new models arrive - always read them with a date attached.

Vision

Multimodal scores

Headline figures use the maximum 1120-token budget. The current default is 280 - the second table shows what that costs.

Benchmark · 1120 tokens	31B	26B A4B	12B	E4B	E2B	Gemma 3 27B
MMMU Pro	76.9	73.8	69.1	52.6	44.2	49.7
MATH-Vision	85.6	82.4	79.7	59.5	52.4	46.0
InfographicVQA	92.0	89.3	88.4	70.0	63.9	70.6
MedXpertQA MM	61.3	58.1	48.7	28.7	23.5	-
OmniDocBench 1.5 lower is better ↓	0.131	0.149	0.164	0.181	0.290	0.365

Benchmark · 280 tokens default	31B	26B A4B	12B	E4B	E2B
MMMU Pro	75.8	73.2	67.7	51.4	43.2
MATH-Vision	83.4	80.3	76.7	59.2	53.0
InfographicVQA	82.8	77.8	58.7	54.8	44.6
MedXpertQA MM	60.7	55.7	47.4	28.7	22.5

Document-heavy tasks are the most sensitive: InfographicVQA falls from 92.0 to 82.8 on the 31B, and from 88.4 to 58.7 on the 12B.

Long context

Retrieval and reasoning across the window

Measured without thinking mode.

Benchmark	Length	31B	26B A4B	12B	E4B	E2B	Gemma 3 27B
RULER	32K	96.8	97.3	96.4	95.2	83.0	91.1
RULER	128K	96.4	89.8	91.2	86.6	70.4	66.0
LOFT Recall@k	128K	79.5	66.3	66.4	58.5	50.5	8.6
GraphWalks F1	<128K	82.3	72.6	71.0	50.9	4.1	32.8
MTOB eng→kgv	128K	52.9	50.0	45.1	37.8	15.4	41.0
MTOB eng→kgv	256K	54.3	48.9	41.9	-	-	-

Audio

Transcription and speech translation

Benchmark	12B	E4B	E2B	3n E4B	3n E2B
FLEURS word error rate, lower better ↓	0.063	0.075	0.090	0.085	0.108
CoVoST CorpusBLEU into English	42.3	38.2	35.4	34.7	31.6

Against Gemma 3n at matching sizes: translation improved 12% (E2B) and 10% (E4B), transcription 17% and 12% - while the audio encoder shrank from 680M to 305M parameters (390 MB → 87 MB on disk). The 31B and 26B A4B accept no audio at all.

Serving

Throughput and latency

vLLM on H100 80GB. Independent measurement, not vendor-reported.

Measurement	31B dense	26B A4B MoE	Conditions
Decode, baseline	40.3 t/s	177.1 t/s	Concurrency 1
Decode, with MTP	125.3 t/s	264.2 t/s	Concurrency 1
MTP speedup	3.11×	1.49×	Dense gains more - it was more bandwidth-bound
Aggregate throughput	1,260 t/s	-	1× H100, ISL 512 / OSL 256
Aggregate, 8× H100	2,208 t/s	-	1.75× for 8× hardware - sublinear
Time to first token	279 ms	-	1× H100, concurrency 1, 128-token prompt
TTFT, 8× H100	85 ms	-	Same prompt
TTFT stable to	~concurrency 8	-	Climbs after, as the prefill queue builds
Weights resident	62 GB	-	BF16, leaving ~18 GB of an 80 GB card for KV cache

FlashAttention 4 on Hopper GPUs, shipped July 2026: prefill throughput +25–70%, time-to-first-token down as much as 31%.

On-device

Mobile performance

LiteRT-LM builds. The GPU-versus-CPU gap is the single largest configuration effect on this page.

Device · backend	Model	Prefill	Decode	First token	Memory
Galaxy S26 Ultra · GPU	E2B	3,808 t/s	52.1 t/s	0.3 s	676 MB
Galaxy S26 Ultra · CPU	E2B	557 t/s	46.9 t/s	1.8 s	1,733 MB
Galaxy S26 Ultra · GPU	E4B	1,293 t/s	22.1 t/s	0.8 s	710 MB
Galaxy S26 Ultra · CPU	E4B	195 t/s	17.7 t/s	5.3 s	3,283 MB
iPhone 17 Pro · GPU	E4B	1,189 t/s	25.1 t/s	0.9 s	-
Dragonwing IQ8 · NPU	E2B	3,700 t/s	31 t/s	-	-
Raspberry Pi 5 · CPU	E2B	133 t/s	7.6 t/s	-	-

Mobile file sizes: E2B 2.58 GB (web variant 2.0 GB, Qualcomm NPU build 2.97 GB); E4B 3.66 GB (2.24 GB decoder plus 0.67 GB memory-mapped embeddings).

Distribution

Checkpoints and licensing

Each suffix explained on the variants page.

Checkpoint	Format	Available for	Purpose
`-it`	Safetensors BF16	All five	Instruction-tuned - the default
no suffix	Safetensors BF16	All five	Base, pre-trained only
`-it-qat-q4_0-gguf`	GGUF 4-bit	All five	Quantisation-aware trained 4-bit
`-it-assistant`	Safetensors	All five	MTP drafter, ~0.4B - not a chat model
`-it-qat-q4_0-unquantized`	Safetensors	Some sizes	QAT weights at full precision
`-it-litert-lm`	.litertlm	E2B, E4B only	Mobile, via `litert-community/`

Release	Date	Contents
Core release	March 31, 2026	E2B, E4B, 26B A4B, 31B
Public announcement	April 2, 2026	-
MTP checkpoints	April 16, 2026	E2B, E4B, 26B A4B, 31B drafters
12B Unified	June 3, 2026	Encoder-free multimodal
Weights refresh	July 15, 2026	FlashAttention 4, tool-calling fixes, vision defaults, chat template - no version bump

Licence: Apache 2.0 for the entire Gemma 4 family. Gemma 3 and earlier, and most specialised variants including ShieldGemma, remain under the Gemma Terms of Use.

Provenance

Where each number comes from

Worth knowing which figures are vendor-reported and which are independent, because they don't tell the same story.

Data	Source	Type	Caveat
Architecture, memory, benchmarks	Gemma 4 Technical Report	Self-reported	Thinking mode on; vision at 1120 tokens
Modalities, context, model IDs	Google model documentation	Self-reported	-
LMArena Elo	LMArena	Independent	Ranks decay as new models arrive - dated June 19, 2026
Throughput, TTFT, MTP speedups	Third-party vLLM benchmarking	Independent	One workload shape; yours will differ
On-device performance	LiteRT community builds	Independent	Flagship hardware - treat as a ceiling
Agentic and SWE-Rebench results	Third-party aggregation	Independent	Substantially less flattering - see limitations

📅

Two things that quietly invalidate comparisons. The July 15, 2026 refresh changed weights without changing the version number, so results measured before and after aren't strictly comparable. And every benchmark here assumes thinking mode enabled and vision at maximum resolution - neither of which is the default configuration you'll get out of the box.

Numbers are one view.

See what the architecture actually does, or where the measurements stop being flattering.

Architecture explained Benchmarks in context Limitations