Skip to content
Search models or paste a model link. Press Enter to search.

Gemma 4 12B

google/gemma-4-12B-it
Left and right arrow keys switch the vendor.
google/gemma-4-12B-it architecture graph
google/gemma-4-12B-it architecture graph
  • huggingface
  • Ryzen AI Max+ 395
Labels
  • Image
  • Text
  • VLM
  • Generation
Build
Embedl-optimized
Precision
W4A4
Runtime
llama.cpp with Embedl optimizations
Context length
8,192 tokens
Run Gemma 4 12B on Ryzen AI Max+ 395

Google's mid-size Gemma 4 model, for chat, reasoning and coding, with images.

Methodology

  • Throughput: one user, real conversations from the public OpenAssistant set, 2,048 tokens in, 128 out, greedy, no prompt reuse. Average of 20 requests after 3 warm-up requests. Baseline llama.cpp runs with its default server settings. End-to-end is the whole request.
  • Power and memory: sampled on the device during those runs. Idle is the device at rest before the run; energy per token is the run's energy over its generated tokens; peak memory is the model server's.
  • Accuracy: full test sets, greedy, the same model files and settings as served. GSM8K 1319, IFEval 541, MMLU-Pro 1008 questions.
  • Baseline llama.cpp: the published weights on upstream llama.cpp, same settings, same device.
  • Reproduce: embedl bench google/gemma-4-12B-it/ryzen-ai-max-395 on your device.

Files

Image Size

Model

docker.io/embedl/model-artifacts:gemma4-12b.huggingface_int4_embedl_llamacpp-rocm.0c4a41c593cba53a

10.9 GiB

Runtime

docker.io/embedl/models:llamacpp-rocm_amd64_20261012_1040d4ee

0.35 GiB

Variants

Every build of Gemma 4 12B on Ryzen AI Max+ 395.

Variant Precision Speed Peak memory Energy / token Runtime

huggingface_int4_embedl_llamacpp-rocm

W4A4 27.7 tok/s
Throughput: End-to-end
14.2 GiB 4.28 J
llama.cpp with Embedl optimizations
llamacpp-rocm_amd64_20261012_1040d4ee

huggingface_int4_llamacpp_llamacpp-rocm

Q4_0 9.9 tok/s
Throughput: End-to-end
16.6 GiB 6.99 J
llama.cpp (Vulkan)
llamacpp-rocm_amd64_20261010_23204e6d