Gemma 4 Download Guide: Every Model & Format (2026)

Google's Gemma 4 comes in five sizes, three usable formats, and — as of June 2026 — an extra set of official QAT builds. That's a lot of names to keep straight when all you want is "give me the file." This page is the hub: every repo, every quant, every file size, verified directly against Hugging Face and Ollama's listings, plus the license question people keep getting wrong.

Quick Answer: Fastest Route by User Type

If you're not sure which of the five sizes to grab, jump to the decision helper below — it's keyed to your available RAM/VRAM.

The Five Models, and Where to Get Them

Gemma 4 ships as E2B, E4B, 12B, 26B A4B, and 31B. Every size is multimodal (text + image, with audio on the three smaller ones), and every size is distributed in at least three formats: raw safetensors on Hugging Face, GGUF for llama.cpp-based tools (Ollama, LM Studio, llama.cpp itself), and MLX for Apple Silicon. Google also shipped official quantization-aware-training (QAT) checkpoints on June 5, 2026, for four of the five sizes — more on the exception below.

Master Download Table

All repos and file sizes below were checked directly against their Hugging Face and Ollama listings.

Each GGUF repo also ships Q6_K and Q8_0 quants (and a small mmproj file needed for image input) if Q4_K_M isn't precise enough for your use case. Unsloth mirrors every size too (e.g. unsloth/gemma-4-12b-it-GGUF), with a wider spread of quant levels down to 2-bit — worth a look if Q4_K_M is still too big for your machine.

Official QAT Checkpoints (June 5, 2026)

Google trained dedicated quantization-aware checkpoints and released them as -qat-w4a16-ct repos, built in the compressed-tensors format for high-throughput serving with vLLM rather than for llama.cpp-based apps:

26B A4B is the exception — but only in this vLLM format. Google didn't ship a w4a16-ct build for it: the MoE's expert dimension (704) is too small for native 4-bit QAT to hold quality in that format, and Google recommends INT8 for vLLM serving of 26B A4B instead.

For llama.cpp-based apps (Ollama, LM Studio), Google did ship official Q4_0 GGUF QAT checkpoints for all five sizes — including google/gemma-4-26B-A4B-it-qat-q4_0-gguf — with one caveat: independent testing by Unsloth found the 26B A4B plain-Q4_0 conversion measurably hurts accuracy, so for that one model a mixed-precision community quant (e.g. Unsloth's dynamic GGUF) is the safer 4-bit choice. On Ollama the QAT builds are one command away via per-size tags like gemma4:12b-it-qat (7.2 GB) or gemma4:31b-it-qat (19 GB); LM Studio packages the 12B one as lmstudio-community/gemma-4-12B-it-QAT-GGUF (Q4_0, 6.98 GB).

Kaggle and ai.google.dev

If you'd rather not use Hugging Face, Google also mirrors every Gemma 4 variant on Kaggle Models, which is convenient if you're already working in a Kaggle notebook with free GPU access. ai.google.dev/gemma/docs is the canonical documentation hub and links out to both Hugging Face and Kaggle for every release.

Is Gemma 4 Free? The License, Explained

Yes — and this is the one thing worth getting exactly right. Every Gemma 4 model, at every size, ships under a plain Apache 2.0 license. That's confirmed on the individual Hugging Face model cards (E2B and 12B both show license: apache-2.0) and in Google's own announcement: "The release of Gemma 4 under the Apache 2.0 license — our most capable open models ranging from edge devices to 31B parameters."

This is a real change from earlier Gemma generations, which shipped under Google's custom "Gemma Terms of Use" — permissive for most uses, but not an OSI-approved license, and it carried content and field-of-use restrictions that made some legal teams nervous. Gemma 4 drops that entirely.

What Apache 2.0 means in practice:

• Commercial use is unrestricted — build and sell a product on top of it, including one that competes directly with Google.
• No requirement to open-source your modifications. Fine-tune it, keep the weights private, ship it in a closed product.
• No per-industry or per-use-case carve-outs, unlike the old Gemma Terms of Use.
• You do need to preserve the license and copyright notice if you redistribute the model itself (standard Apache 2.0 attribution, not a restriction on your product).

Every quant and format on this page — GGUF, MLX, the official QAT checkpoints — inherits the same Apache 2.0 license from the base weights. There's no size or format that's more or less "free" than another.

Which File Should You Download?

The right file depends entirely on your hardware, not on which one sounds the most impressive. We cover this in full detail — with exact VRAM/RAM numbers for every model and quant — in the Gemma 4 Hardware Requirements guide. The short version:

On a Mac, prefer the MLX build over GGUF when one exists for your size — it talks to Metal directly and typically runs 10–20% faster than the GGUF/llama.cpp path on the same hardware. For a device-specific breakdown, see our Gemma 4 on Mac Mini guide.

Per-Tool Quickstarts

Ollama (terminal, fastest)

ollama pull gemma4:12b
ollama run gemma4:12b

Swap 12b for e2b, e4b, 26b, or 31b depending on your hardware. Full walkthrough, including the multimodal API: Run Gemma 4 with Ollama.

LM Studio (GUI)

Press ⌘/Ctrl + Shift + M, search "Gemma 4," and pick the size LM Studio recommends for your machine — it detects your RAM/VRAM automatically and will steer 16 GB Macs toward 12B in GGUF or MLX. Full walkthrough: Run Gemma 4 with LM Studio.

Hugging Face (manual, for developers)

pip install huggingface_hub
huggingface-cli download google/gemma-4-12B-it --local-dir ./gemma-4-12b

Or for a single GGUF file instead of the whole safetensors repo:

huggingface-cli download lmstudio-community/gemma-4-12B-it-GGUF \
  gemma-4-12B-it-Q4_K_M.gguf --local-dir ./gguf

This is the route to take if you're fine-tuning, running through transformers/mlx-lm directly, or need a quant level not covered by the GUI tools.

FAQ

Is Gemma 4 free? Yes, completely — Apache 2.0 across every model size and every format, including commercial use. No account, API key, or payment required to download or run it locally.

GGUF vs. safetensors — which do I want? Safetensors is the original, full-precision (or near-full-precision) format used by transformers, vLLM, and fine-tuning frameworks. GGUF is a quantized, single-file format built for llama.cpp-based tools (Ollama, LM Studio, llama.cpp itself) — smaller downloads, faster CPU/consumer-GPU inference. If you're just chatting locally, get GGUF. If you're fine-tuning or serving at scale, get safetensors.

Which quant should I get? Q4_K_M is the default recommendation for almost everyone — it's the smallest quant that doesn't noticeably degrade output quality. Only step up to Q6_K or Q8_0 if you have VRAM to spare and want to squeeze out the last bit of quality; step down to a QAT build (12B) or Unsloth's lower-bit quants if you're memory-constrained.

What's the difference between the official QAT build and a regular GGUF quant? Both compress the model to roughly 4 bits per weight. The difference is when the compression happens: QAT (quantization-aware training) bakes the low-precision behavior in during training, so quality holds up better at low bit depths. A regular GGUF quant is post-training — Google or the community compresses an already-trained model afterward. QAT is the better pick when it's available for your size; right now that's E2B, E4B, 12B, and 31B, not 26B A4B.

Do I need to convert anything myself? No, for any of the models and quants listed on this page. Conversion only matters if you fine-tune the safetensors weights yourself and want to run the result through Ollama or LM Studio — see the Ollama guide for that conversion path.

Where do I go next? Confirm your hardware fits the model you picked in the Hardware Requirements guide, then follow either the Ollama or LM Studio quickstart above to get it running.