// Gemma · Omni / audio LLM
gemma-4-12B-it-qat-q4_0-unquantized: VRAM & local deployment
12B parameters. About 9.3 GB at 4-bit and 27.7 GB at BF16 with one 8K-token request.
released 2026-06-04licence: Commercial OKmemory figures are estimates
01Specs
| Publisher | Google · google |
|---|---|
| Family / series | Gemma · gemma-4-qat-q4_0-unquantized |
| Type | Omni / audio LLM (any-to-any) |
| Parameters | 12B Source: counted from official safetensors weights |
| Context length | 256K tokens (262,144)max_position_embeddings in config.json |
| Architecture | 48 layers (8 full-attention) · 40 sliding-window layers (1024) · GQA |
| Released weights | BF16 · 23.9 GBdownload |
| Licence | Apache-2.0 · Commercial OKPermissive licence (Apache-2.0 / MIT / CC-BY / OpenMDW): commercial use allowed. |
| Released | 2026-06-04 (Hugging Face repo created) |
02How much VRAM it needs (estimate)
Total = weights + 10% + 0.5GB overhead + KV cache for one request (FP16 KV). More concurrent users multiply the KV column.
| Precision | Weights | Total @ 8K | Total @ 32K | Total @ 128K | Smallest GPU | Mac |
|---|---|---|---|---|---|---|
| BF16 / FP16 (original) | 23.9 GB | 27.7 GB | 29.3 GB | 35.7 GB | RTX 5090 32GB | Mac 64GB |
| FP8 / INT8 / Q8_0 | 12.7 GB | 15.4 GB | 17.0 GB | 23.4 GB | RTX 4090 24GB | Mac 36GB |
| 4-bit (Q4_K_M / AWQ / GPTQ-Int4) | 7.3 GB | 9.3 GB | 11.0 GB | 17.4 GB | 12GB GPU | Mac 16GB |
KV cache per request: 8K 0.9 GB · 32K 2.5 GB · 128K 8.9 GB
Bits per weight: BF16 16, FP8 8.5, 4-bit 4.85. “Fits” = total ≤ 90% of memory; Mac GPU can use ~2/3 of unified memory (≤36GB) or ~3/4 (larger). For MoE, all experts must be in memory, so size by total parameters. Method →
03How to run it
Only runtimes the publisher or the runtime’s official library names for this model.
- vLLM — mentioned in the official model card
- llama.cpp — mentioned in the official model card
- Transformers — mentioned in the official model card
04Official repositories & quantised builds
3 official repo(s) from google.
- gemma-4-12B-it-qat-w4a16-ct Instruct/W4A16
- gemma-4-12B-it-qat-q4_0-gguf Instruct/GGUF
- gemma-4-12B-it-qat-q4_0-unquantized Instruct
05Other Gemma models
All 48 →| Model | Params | 4-bit total | Released |
|---|---|---|---|
| gemma-4-E2B-it-qat-q4_0-unquantized · same series | 5.1B | 4.0 GB | 2026-04-29 |
| gemma-4-E4B-it-qat-q4_0-unquantized · same series | 7.94B | 5.9 GB | 2026-04-30 |
| gemma-4-26B-A4B-it-qat-q4_0-unquantized · same series | 25.8B / 4B | 18.3 GB | 2026-04-29 |
| gemma-4-31B-it-qat-q4_0-unquantized · same series | 31.3B | 23.5 GB | 2026-04-28 |
| embeddinggemma-2 | 744M | 1.0 GB | 2026-09-14 |
| diffusiongemma-26B-A4B-it | 25.8B / 4B | 18.3 GB | 2026-06-09 |
| gemma-4-12B-it | 12B | 9.3 GB | 2026-05-23 |
| gemma-4-26B-A4B-it | 25.8B / 4B | 18.3 GB | 2026-03-11 |
| gemma-4-31B-it | 31.3B | 23.5 GB | 2026-03-11 |
| gemma-4-E2B-it | 5.12B | 4.0 GB | 2026-03-02 |
| gemma-4-E4B-it | 8B | 6.0 GB | 2026-03-02 |
| translategemma-4b-it | 4.3B | 3.4 GB | 2026-01-12 |
06Sources
Data checked 2026-07-20. Parameters, context and licence are copied from the official repository; memory figures are our estimates.