Gemma 4 12B: Inside Google's Encoder-Free 12B Model

Gemma 4 12B is the mid-size member of the Gemma 4 family from DeepMind: a dense, 11.95-billion-parameter open-weights model that reads text, images, audio and video frames and writes text, and whose defining engineering decision is that it carries no multimodal encoders at all. It sits between the edge-focused Gemma 4 E4B and the larger, sparser Gemma 4 26B A4B. The launch note frames its job as bringing agentic multimodal work to a laptop — reasoning, tool use, coding and document or screen understanding that runs on the machine in front of you rather than over a metered connection. It is also the size that carries this family's entire live footprint on our own site: three comparison pages are written about it, and until this one there was no page about the model itself.

That inversion is worth naming up front, because it explains the shape of everything below. Readers do not arrive at Gemma 4 12B the way they arrive at a flagship. They arrive sideways, usually from a matchup page, often because the real question is which size of Gemma 4 — and they want to know what the left-hand side of the comparison actually is before they spend a download on it. The searches we measure for this name are overwhelmingly comparative; the bare term on its own is thin. So this is a reference page: what the 12B is, what its model card states, what it takes to run, what its licence permits, and which of the three existing comparisons already settles the question you came with.

It is not a launch story, and it should not be read as one. Gemma 4 12B shipped on June 3, 2026 — 132 days before this page — and nothing about the model changed in the seven days before publication. That date comes from the vendor's own release note, and it is used here to place the model in time rather than to announce it.

What the 12B is, and where it differs from the 26B A4B

The Gemma 4 family has five sizes, and they are not five points on one line. The model card lists them as E2B, E4B, 12B Unified, 26B A4B and 31B, and they split across two architectures: three dense models and one mixture-of-experts, with the E-sizes using per-layer embeddings to shrink what actually runs on a device.

Gemma 4 12B is a dense model. The model card gives it 11.95B total parameters across 48 layers, with a 1,024-token sliding window and a 256K-token context. The Gemma 4 26B A4B, by contrast, is the family's mixture-of-experts: 25.2B total parameters but only 3.8B active per token, spread over 30 layers with 8 active experts out of 128 plus one shared expert. That "A" is the whole point of the name — the model carries 26B of weights and pays for roughly 4B of them on each forward pass, which is why the model card describes it as running almost as fast as a 4B model while holding far more knowledge.

So the honest one-line distinction is this. The 26B A4B is the size you pick when you want the most capability per unit of latency and you have a workstation or a server to put it on. The 12B is the size you pick when the machine is a laptop and the budget is 16GB of memory. The launch note frames the 12B as the bridge between the E4B and the 26B A4B, and the spec sheet bears that out.

The more interesting difference is architectural, and it is the reason the 12B is labelled "Unified" rather than just "Dense". Every other multimodal model in the family routes images and audio through dedicated encoders before handing the result to the language model — the card lists a ~150M vision encoder and a ~300M audio encoder on the E-sizes, and a ~550M vision encoder on both the 31B and the 26B A4B. The 12B has neither; the column for vision and audio encoder parameters on the 12B reads as a dash in both cases.

What replaced them is small. The launch post describes vision as a lightweight embedding module — a single matrix multiplication plus positional embedding and normalisation — and audio as removing the encoder entirely and projecting the raw signal into the same dimensional space as text tokens. Raw image patches and audio waveforms go straight into a single decoder-only transformer. The stated payoff is lower multimodal latency and the ability to fine-tune the whole thing in one pass, across modalities, with a single adapter rather than one per encoder.

That is also why the 12B is the first mid-size Gemma with native audio input. On the card, "Text, Image, Audio" is the modality row for E2B, E4B and 12B; the 31B and 26B A4B are marked "Text, Image" only.

The specification, as the model card states it

Everything in this section is from the Gemma 4 model card on ai.google.dev, which covers the whole family rather than the 12B alone. Where the card is silent, that is noted rather than filled in.

• Context length — 256K tokens on the 12B. The card notes the split explicitly: the small models (E2B, E4B) get 128K, the medium models (12B, 26B A4B, 31B) get 256K.

• Modality — text, image and audio input; text output. The card lists video as a family capability processed as frame sequences, and describes image input at variable aspect ratio and resolution across all sizes.

• Total parameters — 11.95B, dense, 48 layers. The card gives no active-parameter figure because there are no experts to activate.

• Sliding window — 1,024 tokens, double the 512 the E-sizes use.

• Vocabulary — 262K, shared across all five sizes.

• Attention — a hybrid that interleaves local sliding-window attention with full global attention, with the final layer always global. Global layers use unified keys and values plus proportional RoPE to hold down memory at long context.

• Languages — pre-trained on 140+ languages, with the card claiming out-of-the-box support for 35+.

Two numbers worth reconciling, because they look like a disagreement and are not. The model card says 256K tokens; Artificial Analysis's model page for Gemma 4 12B says 262k. Those are the same figure written two ways — 262,144 tokens is 256K — and the card's 262K vocabulary figure sitting nearby is what makes the pair look contradictory at a glance.

On benchmarks, the card publishes a family table for the instruction-tuned checkpoints, and these are vendor-reported numbers, not independent ones. For Gemma 4 12B the card lists 77.2% on MMLU Pro, 78.8% on GPQA Diamond, 77.5% on AIME 2026 without tools, 72.0% on LiveCodeBench v6, a Codeforces ELO of 1659, 69.1% on MMMU Pro, and 43.4% on MRCR v2 at 128K with eight needles. The same table puts Gemma 4 26B A4B at 82.6% MMLU Pro, 88.3% AIME, 77.1% LiveCodeBench and 44.1% MRCR — ahead on most rows, which is what you would expect from the bigger model, and close enough on a few that the size gap is not the whole story.

The independent read, from Artificial Analysis, is a single headline: an Intelligence Index score of 14, ranked 20th of 141 models in its comparison class, with a measured output speed of 110.9 tokens per second and a blended cost of $0.10 per million input tokens and $0.30 per million output. Those cost figures are a median across third-party platforms that serve the open weights, not a vendor list price — this is worth being precise about, because the vendor publishes no per-token price for the 12B on the model card at all.

Running it: what it actually needs

The launch post gives the headline requirement directly: the 12B is "small enough to run locally with just 16GB of VRAM or unified memory." The quantisation sizes published for it make that concrete.

• Full precision — the instruction-tuned checkpoint on Hugging Face is a single 23.92 GB safetensors file. At bfloat16 that is the raw weight footprint, before any KV cache, so a 16GB machine cannot hold it and a 24GB card is tight.

• The vendor's own quantised builds — because the 12B shipped with quantisation-aware training rather than only a post-hoc quantiser. The vendor publishes a QAT GGUF at roughly 6.98 GB for the weights plus a 0.18 GB multimodal projection file, and a QAT compressed-tensors build (w4a16) at 10.26 GB. There is also an unquantised QAT checkpoint published specifically as the base to fine-tune from before re-quantising.

• Community GGUF — the widely mirrored Unsloth build runs from about 4.2 GB at the most aggressive 2-bit through 6.38 GB at IQ4_XS, 7.12 GB at Q4_K_M, 8.41 GB at Q5_K_M, and 12.67 GB at Q8_0. The LM Studio mirror publishes Q4_K_M at 7.38 GB, Q6_K at 9.79 GB and Q8_0 at 12.67 GB, each with the same 0.18 GB projection file.

• MLX — an 8-bit quantised build for Apple silicon lands around 12.7 GB across three files; a 4-bit community build of a coding fine-tune comes in near 6.7 GB.

Read together, that is the practical answer to "will this fit." A 4-bit build of Gemma 4 12B around 7 GB is the version that makes a 16GB laptop work, exactly as the vendor claims. If you want the model unquantised, the target is a 24GB card or more, and if you want the full 256K context you need headroom on top of the weight footprint for the cache.

Two things the sources do not give you, and which this page will not invent. There is no throughput figure published by the vendor for local inference — the 110.9 tokens per second from Artificial Analysis is a hosted measurement on third-party infrastructure, not a laptop number, and the two are not interchangeable. And there is no published minimum-RAM table per quantisation; the sizes above are file sizes, which is a floor, not a requirement.

What the model card does document is the surrounding toolchain rather than any single blessed runtime: Hugging Face Transformers, llama.cpp, MLX, SGLang, vLLM and Unsloth for inference and fine-tuning, with LM Studio, Ollama and LiteRT-LM as the packaged local paths, and the weights on Hugging Face and Kaggle. The 12B also ships with multi-token-prediction drafters, published as separate assistant checkpoints — a small additional download that trades memory for latency.

One routing fact, stated plainly because it is the kind of thing a comparison page leaves ambiguous: Gemma 4 12B is not a model OrcaRouter serves, on any spelling — not the bare name, not a dated variant, not a quantised one. The Gemma 4 sizes we do route are the 26B A4B and the 31B. So if you want to call this model, self-hosting the weights or a third-party host that serves them is the route; if you would rather not run it yourself, the honest next step is to check whether a different size of the family gets you what you needed.

The licence, for this size specifically

The Gemma 4 model card states the licence as Apache 2.0 and links to the Gemma 4 licence page, which serves the Apache 2.0 text — the standard Apache licence, with no usage caps, no monthly-active-user threshold and no separate commercial tier. The Hugging Face repository for the 12B instruction-tuned checkpoint carries the same declaration in its metadata and points at the same licence URL. So for this size the licence question has a short answer: use it, modify it, redistribute it, commercially or otherwise, under Apache 2.0 terms.

Two collateral documents exist alongside it and are worth knowing about rather than confusing with the licence. The vendor publishes a Gemma Terms of Use page and a Gemma Prohibited Use Policy. The prohibited-use policy is a behavioural document listing categories of output the models must not be used to generate. It is a policy, not a licence restriction on the weights — the weights themselves are Apache 2.0, which is what makes community derivatives of this model possible at all. If you are redistributing Gemma 4 12B or a fine-tune of it, the licence is what governs your redistribution; the policy is what governs acceptable use.

That distinction answers the question the community derivative ecosystem raises, without needing to characterise anyone's behaviour. Because the weights are open under a permissive licence, third parties can and do publish fine-tunes and quantisations of the 12B — abliterated variants, domain fine-tunes, alternate quantisation recipes. What the licence permits is the redistribution of those derivatives; the licence is also why they exist.

The three comparisons this page sits behind

All three live pages under this stem are matchups, and each one already answers a different question. This page does not adjudicate any of them — it is the thing they describe.

• Gemma 4 12B vs Gemini 3.5 Flash-Lite — settles what happens when both options are the vendor's and the axis is how you pay: a model you download under Apache 2.0 against the cheapest metered rung of the hosted API.

• Gemma 4 12B vs GLM 5.2 — settles the wildly asymmetric pairing: an encoder-free 12B you point at a laptop versus a 744B-parameter open-weight mixture-of-experts built for long-horizon agentic coding at a large multiple of the price.

• Gemma 4 12B vs Qwen3.8 Max — settles the on-device-versus-rented question directly: a machine on your desk against the same lab's flagship served by the token.

If what you actually want is one of the other Gemma 4 sizes rather than another 12B matchup, the page that covers all five at once is the family hub, which is where the E2B, E4B, 26B A4B and 31B are described together. And if you have already decided the 12B is not the right size, the models it is most often weighed against — GLM 5.2, Qwen3.8 Max and Gemini 3.5 Flash-Lite — are all reachable through OrcaRouter on one key, with 0% markup on the provider's own list price, so a vendor price cut lands here the same day it is announced.

What is not settled

Two gaps are worth stating rather than papering over. The benchmark table above is vendor-reported — the vendor's own card and evaluation setup — and while Artificial Analysis has since scored the model independently at an Intelligence Index of 14, there is no broad independent replication of the per-benchmark figures, so treat the row-level numbers as claims. And there is no published per-quantisation memory table for the 12B, which means the "16GB is enough" line is a vendor claim with file sizes that support it, not a measured configuration you can copy.

The model itself is easy to place: an Apache 2.0, encoder-free, 12B-parameter multimodal model, dated to June 3, 2026, that will run on a laptop at 4-bit and is the smallest Gemma 4 size with native audio. What it is not is a model with a hosted first-party rate, which is the reason the practical decision around it is a hardware decision first and a price decision second.

If you have landed here from a matchup and the matchup is not the one you needed, the next page is the comparison that matches your question, and failing that the wider Gemma 4 family — the 12B is one size of five, and the size you actually want may be the 26B A4B you can call today or the 31B you can run at the top of the family's range.