Gemma 4 & EmbeddingGemma 2: Model Variants & Specifications
Five Gemma 4 sizes, nine kinds of download, and one new search model in four configurations. This page lists every variant with its specifications, and decodes the names so you download the right thing first time.
Checked against Google's model cards, Hugging Face collections and Ollama's
Five sizes at a glance
Every checkpoint type below exists in some or all of these five sizes. Pick the size first, then the format. For what each can do see features and benchmarks; the mid-size model has its own Gemma 4 12B guide.
| Model | Architecture | Parameters | Layers | Context | Inputs | Drafter |
|---|---|---|---|---|---|---|
| Gemma 4 31B | Dense | 30.7B | 60 | 256K | Text, image | 0.5B |
| Gemma 4 26B A4B | Mixture of Experts | 25.2B (3.8B active) | 30 | 256K | Text, image | 0.4B |
| Gemma 4 12B | Unified, encoder-free | 11.95B | 48 | 256K | Text, image, audio | 0.4B |
| Gemma 4 E4B | Edge | 4.5B effective (8B) | 42 | 128K | Text, image, audio | 79M |
| Gemma 4 E2B | Edge | 2.3B effective (5.1B) | 35 | 128K | Text, image, audio | 78M |
From Google's Gemma 4 model card; drafter sizes as listed on Hugging Face. All five sizes can process video as a sequence of image frames; audio input is limited to E2B, E4B and 12B. Figures in brackets for the edge models include their embedding tables. Memory needs are in the VRAM guide.
Gemma 4 spec sheet
The numbers for all five sizes in one table.
| Specification | 31B | 26B A4B | 12B | E4B | E2B |
|---|---|---|---|---|---|
| Design | Dense | Mixture of Experts | Unified, encoder-free | Edge | Edge |
| Parameters | 30.7B | 25.2B | 11.95B | 4.5B effective (8B) | 2.3B effective (5.1B) |
| Active per token | All | About 3.8B | All | All | All |
| Layers | 60 | 30 | 48 | 42 | 35 |
| Context window | 256K | 256K | 256K | 128K | 128K |
| Text and image input | Yes | Yes | Yes | Yes | Yes |
| Video, as frames | Yes | Yes | Yes | Yes | Yes |
| Audio input | - | - | Yes | Yes | Yes |
| Vision path | 550M encoder | 550M encoder | 35M projection | 150M encoder | 150M encoder |
| Audio path | - | - | Raw projection | 305M encoder | 305M encoder |
| Drafter size | 0.5B | 0.43B | 0.4B | About 78M | About 78M |
| Memory, full precision | 69.9 GB | 57.7 GB | 26.7 GB | 17.9 GB | 11.4 GB |
| Memory, 8-bit | 34.9 GB | 28.8 GB | 13.4 GB | 8.9 GB | 5.7 GB |
| Memory, 4-bit | 17.5 GB | 14.4 GB | 6.7 GB | 4.5 GB | 2.9 GB |
| Mobile build memory | - | - | - | 2.5 GB | 1.1 GB |
| Ollama QAT download | 19 GB | 16 GB | 7.2 GB | 6.1 GB | 4.3 GB |
| Output | Text only | ||||
| Thinking mode and function calling | Every size | ||||
| Image detail | 70, 140, 280, 560 or 1,120 tokens per image; 280 by default | ||||
| Audio and video limits | 30 seconds of audio; 60 seconds of video at one frame a second | ||||
| Languages | 140+, with a vocabulary of about 262,000 entries | ||||
| Recommended sampling | Temperature 1.0, top-p 0.95, top-k 64 | ||||
| Training data cutoff | January 2025 | ||||
| Released | March 31, 2026; the 12B on June 3, 2026 | ||||
| Licence | Apache 2.0 | ||||
Parameters, layers, memory, limits and sampling from Google's model card; vision and audio paths from the technical report; drafter sizes as listed on Hugging Face; downloads from Ollama's tag list. Figures in brackets for the edge models include their embedding tables. How it is built is explained on the features page; the datasheet with benchmarks is on technical specs.
Nine checkpoint types
Shown here for the 12B. Three of these arrived with Google's QAT release and were missing from earlier versions of this page.
gemma-4-12BBase model, for your own post-training-itInstruction-tuned - what you want-it-assistantMTP drafter - not a chat model-it-qat-q4_0-ggufOfficial 4-bit GGUF-it-qat-q4_0-unquantizedQAT weights at half precision-it-qat-q4_0-unquantized-assistantDrafter matched to the QAT model-it-qat-w4a16-ctCompressed tensors for vLLM and SGLang-it-qat-mobile-*Mobile-optimised QAT - E2B and E4B only-it-litert-lmLiteRT-LM mobile build - E2B and E4B onlyWhat each part means
Repository names are precise. Every segment carries information.
- googleThe publisher. Anything else is a community re-upload: often fine, occasionally stale or quantised more aggressively than advertised.
- gemma-4The generation. Gemma 3 uses
gemma-3with a lowercase size suffix, which is a real trap when scripting across both. - 12BThe size, with a capital B on Gemma 4 (and capital
EandA4B). Hugging Face currently resolves the lowercase spelling too, but other tools and mirrors are case-sensitive - copy the canonical form. - itInstruction-tuned. No
-itmeans the base pre-trained model, which does not follow instructions. - qat-q4_0-ggufThe format and precision. Absent means full-precision Safetensors.
-it-assistant is not an assistant
The most misleading name in the family. These are Multi-Token Prediction drafters for speculative decoding - not chat models, and not usable on their own.
A tiny draft model
A 4-layer drafter that proposes several tokens ahead for the full model to verify in parallel. About 78M parameters for E2B and E4B, 0.4B for 12B and 26B A4B, and 0.5B for 31B.
Faster decoding, same output
Google describes it as significantly faster inference with no quality loss: the full model still verifies every token it emits. Every size, including 12B, now has one.
Load it alongside the main model
It is a companion, never a replacement. Loading one on its own and expecting a chatbot is the mistake the name invites.
# Load both models with from_pretrained, then pass the drafter to generate() # target: google/gemma-4-12B-it # drafter: google/gemma-4-12B-it-assistant outputs = target_model.generate( **inputs, assistant_model=assistant_model, # the drafter max_new_tokens=256, do_sample=False, ) # Without assistant_model the main checkpoint still works - # just slower. This is free speed most people never claim.
Generation call from Google's MTP guide for Transformers, which has the full loading code. Serving at scale? See self-hosting. Overview: Multi-Token Prediction.
-qat-q4_0-unquantized-assistant drafter, not the standard one.Every checkpoint, every size
Prefix all Hugging Face rows with google/, except the LiteRT builds, which come from litert-community/.
| Checkpoint type | 31B | 26B A4B | 12B | E4B | E2B | Use it when |
|---|---|---|---|---|---|---|
-itSafetensors, full precision | ✓ | ✓ | ✓ | ✓ | ✓ | The default. Instruction-tuned, for serving and fine-tuning |
baseno suffix | ✓ | ✓ | ✓ | ✓ | ✓ | Your own post-training. It will not follow instructions as shipped |
-it-qat-q4_0-ggufofficial 4-bit GGUF | ✓ | ✓ | ✓ | ✓ | ✓ | Running locally on llama.cpp, Ollama or LM Studio - the best 4-bit quality available |
-it-assistantMTP drafter | ✓ | ✓ | ✓ | ✓ | ✓ | Alongside the matching -it model, for faster decoding |
-it-qat-q4_0-unquantizedQAT weights, half precision | ✓ | ✓ | ✓ | ✓ | ✓ | Applying your own quantisation, or converting to another format such as MLX |
-it-qat-q4_0-unquantized-assistantNEWdrafter for the QAT model | ✓ | ✓ | ✓ | ✓ | ✓ | Speculative decoding when the main model is the QAT build |
-it-qat-w4a16-ctNEWcompressed-tensors, 4-bit weights | ✓ | - | ✓ | ✓ | ✓ | High-throughput serving on vLLM or SGLang. No 26B A4B build |
-it-qat-mobile-transformersNEWmobile-optimised QAT | - | - | - | ✓ | ✓ | Reference weights for phone-class deployment |
-it-litert-lmlitert-community/ | - | - | - | ✓ | ✓ | Android or iOS on-device deployment with LiteRT-LM |
Source: Google's Gemma 4, QAT Q4_0 and QAT mobile collections, which remain the authoritative lists. The LiteRT row is carried over from the previous version of this page.
Most people need two
The -it-qat-q4_0-gguf build to run locally, or the plain -it to serve at full precision. Add the matching drafter if you want the speedup.
Base models disappoint
Downloading the no-suffix repo and finding it rambles rather than answering is a common first mistake. That is expected - it has not been instruction-tuned. Use -it.
Mobile has two routes
Ready-to-run .litertlm builds under litert-community/, and Google's own -qat-mobile reference weights. Both exist only for E2B and E4B.
Ollama tags don't match Hugging Face
Same models, different conventions. Translating between them catches people out constantly.
| Hugging Face | Ollama equivalent | Difference to watch |
|---|---|---|
google/gemma-4-12B-it | gemma4:12b | Lowercase b; no hyphen after gemma; -it implied |
google/gemma-4-26B-A4B-it | gemma4:26b-a4b or gemma4:26b | Both tags exist; the short one drops A4B |
google/gemma-4-E2B-it | gemma4:e2b | Lowercase e2b |
...-it-qat-q4_0-gguf | gemma4:12b-it-qat | Shortened; the q4_0-gguf part is implicit |
| no direct equivalent | gemma4:12b-mlx | Apple silicon build, packaged by Ollama |
| no direct equivalent | gemma4:cloud, gemma4:31b-cloud | Runs on Ollama's hosted service, not your machine |
| no direct equivalent | gemma4:latest | Currently the same build as gemma4:e4b - and may be repointed at any time |
Suffixes you will see on every size
| Tag suffix | What it is | Example |
|---|---|---|
| (none) | Ollama's default build for that size | gemma4:31b |
-it-qat | Google's quantization-aware 4-bit build | gemma4:12b-it-qat |
-it-q4_K_M | Standard 4-bit quantisation | gemma4:e4b-it-q4_K_M |
-it-q8_0 | 8-bit - larger, closer to full quality | gemma4:12b-it-q8_0 |
-it-bf16 | Full 16-bit precision | gemma4:e2b-it-bf16 |
-mlx, -mlx-bf16 | Apple silicon (MLX) builds | gemma4:26b-mlx |
-mxfp8, -nvfp4 | Hardware-specific low-precision formats | gemma4:31b-nvfp4 |
From Ollama's gemma4 tag list, 51 tags as of October 7, 2026.
New to Ollama? Follow how to download or Mac setup. Two one-off MTP tags also exist: gemma4:26b-a4b-it-mtp-q4_K_M and gemma4:31b-coding-mtp-bf16.
:latest. It is an alias Ollama can repoint, so the model underneath you
can change without warning. Name the size explicitly - and the quantisation too if it matters.EmbeddingGemma 2 variants and specifications
EmbeddingGemma 2 is built on the Gemma 4 architecture but does a different job: it turns text, code, images, video and audio into vectors for search. It is one model made of three parts, so you load only what you need.
The four configurations
| Configuration | Parameters | Handles | Ollama name | Ollama download | Phone build |
|---|---|---|---|---|---|
| Text and code | 270M | Text, code | embeddinggemma-2:270m | 378 MB | 165 MB |
| Text and vision | 440M | Text, code, images, video | embeddinggemma-2:440m | 714 MB | 388 MB |
| Text and audio | 570M | Text, code, audio | embeddinggemma-2:570m | 990 MB | - |
| Everything | 740M | Text, code, images, video, audio | embeddinggemma-2 | 1.3 GB | 485 MB |
Specifications
| Type | Embedding model; produces vectors, not text |
| Layers | 24 |
| Vector size | 768; can be shortened to 512, 256 or 128 |
| Input budget | 8,192 tokens, shared by everything in one input |
| Fits in one input | About 29 images, 58 video frames or 327 seconds of audio |
| Vocabulary | 262,144 tokens, shared with Gemma 4 |
| Languages | 100+ |
| Precision | bfloat16 or float32; not float16 |
| Licence | Apache 2.0 |
Where each build lives
| Build | Name |
|---|---|
| Official weights | google/embeddinggemma-2 |
| Ollama | embeddinggemma-2, four sizes |
| GGUF, text only | ggml-org/embeddinggemma-2-GGUF, 310 MB or 558 MB |
| Phone, text | litert-community/embeddinggemma-2-text-270m-litert-lm |
| Phone, text and vision | litert-community/embeddinggemma-2-text-vision-440m-litert-lm |
| Phone, full | litert-community/embeddinggemma-2-740m-litert-lm |
| Kaggle | Kaggle model page |
How the names work
Unlike Gemma 4 there is one official repository. The number after the colon on Ollama, or in a phone build's name, tells you which parts are included.
Shorter vectors
Google's card reports little quality loss down to 256 numbers. At 128 the multimodal score falls from 59.01 to 45.65. See limitations.
Not interchangeable with version 1
The original EmbeddingGemma was text only, 308M parameters, on older terms. Vectors from the two versions cannot be mixed.
Configurations, specifications and limits from Google's EmbeddingGemma 2 model card and developer guide; Ollama sizes from its tag list; phone builds from Google's LiteRT-LM overview; the GGUF build is a community conversion. How to install and use it: install · how to use · EmbeddingGemma 2 guide.
Beyond the core five
Two newer open models are built on Gemma 4: EmbeddingGemma 2, covered above, and DiffusionGemma. They are different tools, not bigger or smaller chat models.
Turns text, code, images, video and audio into vectors in one shared 768-dimension space, for search, retrieval and classification. Its variants and specifications are listed above.
- Repository
google/embeddinggemma-2 - Context8,192 tokens
- LicenseApache 2.0
A 26B A4B sibling that writes by refining whole blocks of tokens in parallel instead of one at a time - roughly 4x faster output, with lower benchmark scores. Aimed at in-line editing and code infilling.
- Repository
google/diffusiongemma-26B-A4B-it - Parameters25.2B (3.8B active)
- LicenseApache 2.0
Earlier specialised models - TranslateGemma, MedGemma 1.5, FunctionGemma, ShieldGemma 2 and T5Gemma v2 - predate Gemma 4 and are listed in Google's release history. Comparing against rivals? See the model comparison and Gemma 4 vs Qwen 3.5. For the previous generation see Gemma 3.
Which one should you download?
| What you're doing | Download | Also grab |
|---|---|---|
| Chatting locally | gemma4:12b-it-qat via Ollama | - |
| Serving with vLLM | google/gemma-4-31B-it-qat-w4a16-ct, or google/gemma-4-26B-A4B-it at full precision | The matching drafter for speculative decoding |
| Maximum local speed | google/gemma-4-31B-it | google/gemma-4-31B-it-assistant |
| Fine-tuning for a task | google/gemma-4-12B-it | Base model if you're doing full post-training |
| Custom quantisation or MLX | -it-qat-q4_0-unquantized | QAT weights beat quantising the plain checkpoint |
| Android or iOS app | litert-community/gemma-4-E2B-it-litert-lm | E4B build if the device allows |
| Audio input | google/gemma-4-12B-it | Or E4B / E2B. The 31B and 26B A4B do not accept audio |
| Video input | Any size | Video is processed as image frames; keep clips short |
| Search or RAG retrieval | google/embeddinggemma-2 | A Gemma 4 chat model to write the answer |
Pull commands and links for each are on the download page; to pick a runtime use the runtime selector. Prefer not to download? Try the playground, chat options or a hosted API, and check pricing.
Six ways to download the wrong thing
Lowercase the size
The canonical name is gemma-4-12B-it with a capital B. Gemma 3 used lowercase, which is why the mistake is easy - and case-sensitive tools will not forgive it.
Grab the assistant expecting a chatbot
It is a small speculative-decoding drafter. On its own it does nothing useful - it needs the main model loaded alongside it.
Take a community quant over the QAT build
Google's -qat-q4_0-gguf checkpoints are quantisation-aware trained and lose less quality than post-hoc community quants at the same bit depth.
Download the base model by accident
No -it means no instruction tuning. It will continue your text rather than answer you, which reads like a broken model if you were not expecting it.
Reach below 4-bit
Q2 and Q3 builds load and run, but reasoning and code quality fall off sharply. A smaller model at 4-bit beats a larger one at 3-bit.
Use weights cached before mid-July
The July 2026 refresh fixed the chat template and tool calling without a version bump. Same name, different weights - re-pull.
Variant questions
More on the full FAQ page; usage tips in prompting; known issues in limitations.
What is the -it-assistant variant really for?
Speculative decoding. It is a small Multi-Token Prediction drafter - from about 78M parameters on the edge models to 0.5B on the 31B - that proposes several tokens ahead for the full model to verify in parallel, giving faster decoding at no quality cost.
The name strongly implies a chat model and it isn't one. Load it as a second model alongside the matching -it checkpoint.
Do I need the base model?
Almost certainly not. Base checkpoints have no instruction tuning - they continue text rather than answering
questions. They exist for people doing their own post-training from scratch. For fine-tuning on a task, starting from -it is usually the better choice.
QAT GGUF or a community quant?
The official QAT build - one now exists for every size. Quantisation-aware training exposes the model to rounding error during training so it adapts; post-hoc quantisation just rounds a finished model. Community quants from reputable packagers are fine, but at the same bit depth the official one is generally better. More in the QAT guide.
What's the unquantized QAT variant for?
It holds the quantisation-aware-trained weights at half precision, so you can apply your own quantisation scheme or convert to another format such as MLX and still benefit from QAT. It is published for all five sizes. If you just want 4-bit, take the GGUF.
What are the w4a16-ct builds?
QAT checkpoints saved in the compressed-tensors format with 4-bit weights and 16-bit activations, intended for high-throughput serving on vLLM or SGLang. They exist for E2B, E4B, 12B and 31B - not for the 26B A4B.
Why do Ollama tags look nothing like the repo names?
They are independent naming schemes. Ollama uses gemma4:12b - no hyphen, lowercase, instruction-tuning
implied - while Hugging Face uses google/gemma-4-12B-it. Translate deliberately rather than guessing.
Is gemma4:latest safe to use?
For experimenting, yes. For anything reproducible, no - it is an alias that currently resolves to E4B and can be repointed, changing the model under your application without warning. Pin the size and quantisation explicitly.
Which sizes accept audio and video?
Audio: E2B, E4B and 12B only. Video: all five, because it is processed as a sequence of image frames. Google's model card caps audio at 30 seconds and video at 60 seconds at one frame per second.
Which sizes have mobile builds?
Only E2B and E4B - as .litertlm files under litert-community/, and as Google's
-qat-mobile reference weights. The 12B and larger have no mobile build. See the Android guide
and mobile downloads.
How many EmbeddingGemma 2 variants are there?
One model in four configurations: 270M for text and code, 440M with vision, 570M with audio, and 740M with everything. On Ollama they are separate names; in Python you load the parts you want.
Is EmbeddingGemma 2 a Gemma 4 model?
It is built on the Gemma 4 architecture and shares its tokenizer and audio encoder, but it is a separate model with a different job. It produces vectors for search and cannot chat. See the EmbeddingGemma 2 guide.
Know which file you need?
Exact repository IDs and pull commands for every checkpoint in the family.