Granite-4.1-8B — int4-group ATEX bake for Adam

GPU-resident int4-group (group-128) ATEX bake of IBM's Apache-2.0 granite-4.1-8b, for Adam (texture-native local AI). Standard GQA attention, ~5.9 GB resident, fits a 16 GB GPU with headroom. Decode-on-kernel (int4grp_gemv); codes decode bit-exactly (roundtrip maxerr 0.0) to the quantized weights.

Why this is Adam's default heavy model: ~2.5x faster decode than the prior NVFP4-12B, graph-friendly standard attention, and IBM reports the 8B matches their 32B-A9B MoE on quality. Logicality verified across math/physics/code/logic.

Manual download

from huggingface_hub import snapshot_download
snapshot_download(repo_id='amnibro/granite41-8b-gf17-atex', local_dir='./bakes/granite41_8b_gf17_atex')

License

Inherits Apache 2.0 from upstream Granite 4.1. Adam (the surrounding code) is CC BY-NC 4.0 — see the GitHub repo.

Downloads last month
11
Safetensors
Model size
5B params
Tensor type
F32
·
BF16
·
U8
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for amnibro/granite41-8b-gf17-atex

Finetuned
(27)
this model