gemma-4-12b-it-qat

Model Card for gemma-4-12B-it (QAT, GGUF)

A GGUF repack of Google's quantization-aware-trained gemma-4-12B-it build, packaged to run on device. The tokenizer, the merge table and the chat template are embedded, so nothing ships alongside the file and nothing calls out to a server at inference time.

This repository holds the repacked weights only, not training data or the original checkpoint.

File format

.ggxf is GGUF plus the tensors and metadata keys this project adds, so these files load in Gadeon and are not expected to load in other GGUF readers.

Model Details

Model Description

gemma-4-12B is the "unified" member of the Gemma 4 family, and it is built differently from the E2B / E4B mobile models. There are no per-layer embeddings, no shared key/value layers, and — the part that surprises people — no vision tower and no audio tower. Multimodality is encoder-free: an image becomes 48x48 raw pixel patches through a LayerNorm/Dense pair plus a factorized 2-D position table, and audio is a 640-sample frame of raw 16 kHz waveform through a single projection. Both land directly in the language model's embedding space.

Attention alternates five sliding-window layers to one full-attention layer, and the two kinds differ in more than the window. A sliding layer is 16 query heads over 8 key/value heads at 256 wide with a RoPE base of 10 000. A full layer is 16 query heads over one key/value head at 512 wide, with proportional RoPE at base 1 000 000, and it carries no value projection at all: key and value share one projection, the value branch taking it before the key norm and the rotation.

The upstream checkpoint is quantization-aware TRAINED, not post-hoc quantized.

• Developed by: Google (base model); GGUF repack by leok7v
• Model type: Multimodal (text, image, audio) causal language model
• Language(s): the languages of the base model
• License: Apache 2.0, inherited from the base model; see the Gemma 4 license
• Repacked from model: google/gemma-4-12B-it-qat-q4_0-unquantized

Model Sources

• Repository (this build): https://huggingface.co/leok7v/gemma-4-12b-it-qat
• Base model: google/gemma-4-12B-it-qat-q4_0-unquantized

Uses

Direct Use

On-device assistants and chat in offline or privacy-sensitive settings where prompts must not leave the device.

Out-of-Scope Use

Not for tasks requiring verified factual accuracy, and not a basis for legal, medical, financial, or safety decisions. Not intended for high-stakes automated decisions without human review.

Bias, Risks, and Limitations

A language model predicts plausible text; it does not verify facts and has no understanding of truth. Output can be confident and still be wrong, incomplete, or biased. The base model's biases and limitations carry over; see the base model card.

Recommendations

Treat every response critically and check important output against a trusted source.

How to Get Started with the Model

The tokenizer (262 144 tokens, 514 906 merges, byte fallback and metaspace) and the chat template are embedded in the file, so a runtime reads both from the GGUF rather than from sibling files. The checkpoint's own config.json, generation_config.json, processor_config.json and tokenizer_config.json are carried verbatim as metadata, which is where the reasoning and tool-call parsing contract lives.

Follow the base model card's guidance: sampling at temperature 1.0, top_p 0.95, top_k 64 (these are recorded in the file), thinking enabled by a <|think|> token at the start of the system prompt, image content placed before the text and audio content after it.

Technical Specifications

Model Architecture and Objective

Weight format

6.823 GB, 677 tensors, GGUF v3, 16384-byte tensor alignment.

Repacking, and why it is lossless

-q4_0-unquantized does not mean "not quantized". Training rounded every weight of a 32-wide block onto a 16-level grid w = (q - 8) * d, and the checkpoint then stored the product in bf16. The codes are still in there.

So this repack recovers (q, d) rather than quantizing again. That distinction is measurable: ggml's quantizer picks d from the block's signed extreme as max / -8, which is lossy for any block whose extreme is positive, and would choose fresh codes for weights that were trained where they sit.

Every block is verified individually — a block is written only if q * d, with d already cast to the fp16 a block header stores, reproduces all 32 values to within one bf16 ulp. All 11.8 billion weights passed. A single mis-assigned code would land at least d/2 away, eight times that tolerance, so the check cannot pass a corrupted tensor.

The consequence is that the reconstructed weights are closer to the trained values than the bf16 checkpoint is: the checkpoint holds round_bf16(d * (q - 8)) while this file holds (d, q) and reconstructs in fp32.

Measured against Google's own 4-bit release

Google also publishes gemma-4-12B-it-qat-w4a16-ct, a compressed-tensors int4 build (group 32, symmetric, memoryless_minmax observer) intended for vLLM. It and this file derive from the same q4_0-unquantized checkpoint, so both can be scored against it directly.

Relative weight error against that bf16 checkpoint:

About 65x closer to the trained weights.

That is not a better search. The int4 codes agree 99.40% between the two files, so both recover the same trained grid; w4a16-ct is an independent witness that the recovery is right. The scales are what differ, its own running a median 1.0645x larger: a min/max observer takes d from the block's extreme, where this repack refits d by least squares over the settled codes. Same codes, different d.

w4a16-ct is better on exactly one tensor. It leaves token_embd in bf16 and therefore exact, at 1.88 GiB against 0.53 GiB here, which is most of the difference between its 9.56 GiB and this file's 6.35 GiB.

Both files leave the same modules unquantized: the vision patch dense, both multimodal projections, and the position table.

Citation

Please cite the upstream Gemma model. See the base model card at google/gemma-4-12B-it-qat-q4_0-unquantized for the canonical reference and license.

Model Card Contact

leok7v on Hugging Face.