Table with columns: Tensors, Format| Tensors | Format |
|---|
| attention q, o (42 layers) | bfloat8_b, 8.5 bits a weight |
| attention k, v (24 layers) | bfloat8_b, 8.5 bits a weight |
| MLP gate, up, down (42 layers) | bfloat8_b, 8.5 bits a weight |
| LM head (tied to the embedding), embedding, norms and every other tensor | bf16, as in the base model |
audio_tower, embed_audio, embed_vision, vision_tower | bf16, unchanged |
12.06 bits a weight on average, 11.34 GB packed. quant4.json lists each projection's format. Quantized with
GPTQ on the bfloat8_b grid, keeping each block's shared exponent as the packer sets it, calibrated on 65,536
tokens of WikiText-2 and Korean Wikipedia.
The token and per-layer embeddings stay bf16 (16 bits a weight); the quantized attention and MLP projections use 8.5 bits a weight, so the average over all language-model weights is higher.
Only the language model was quantized and evaluated.
Quickstart
from transformers import AutoModelForCausalLM, AutoTokenizer repo = "Xylolabs/gemma-4-E4B-it-Tenstorrent-Tensix-BFP8"tok = AutoTokenizer.from_pretrained(repo)model = AutoModelForCausalLM.from_pretrained(repo, dtype="bfloat16", device_map="auto") messages = [{"role": "user", "content": "What is the capital of France?"}]inputs = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt", return_dict=True).to(model.device)out = model.generate(**inputs, max_new_tokens=64)print(tok.decode(out[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))
Run on Tensix
Checked with ttnn 0.81.1 on ttsim, Tenstorrent's Blackhole simulator. Pack one matrix and
check that no value changed (TT-NN reads weights as [in, out]):
import jsonimport torchimport ttnnfrom safetensors import safe_open d = "gemma-4-E4B-it-Tenstorrent-Tensix-BFP8" # hf download Xylolabs/gemma-4-E4B-it-Tenstorrent-Tensix-BFP8 --local-dir gemma-4-E4B-it-Tenstorrent-Tensix-BFP8name = "model.language_model.layers.0.mlp.down_proj.weight"fmt = json.load(open(f"{d}/quant4.json"))["layers"]["0"]["mlp.down_proj"]["format"]dtype = {"bfp4": ttnn.bfloat4_b, "bfp8": ttnn.bfloat8_b}[fmt]shard = json.load(open(f"{d}/model.safetensors.index.json"))["weight_map"][name]with safe_open(f"{d}/{shard}", "pt") as f: w = f.get_tensor(name).t().contiguous() device = ttnn.open_device(device_id=0)w_tt = ttnn.as_tensor(w, dtype=dtype, layout=ttnn.TILE_LAYOUT, device=device, memory_config=ttnn.DRAM_MEMORY_CONFIG)assert torch.equal(ttnn.to_torch(w_tt).float(), w.float())ttnn.close_device(device)
python pack_tensix.py gemma-4-E4B-it-Tenstorrent-Tensix-BFP8 checks every matrix this way (--device also round-trips
it through device DRAM). python verify_grid.py gemma-4-E4B-it-Tenstorrent-Tensix-BFP8 checks the grid with torch alone.
tt-metal Gemma 4 demo (v0.81.1). tt-transformers does not support Gemma 4. tt-metal's Gemma 4
demo runs E2B, E4B and 12B on one
Blackhole P150. With no entry for this repository in its precision_overrides.json, every module loads in
bf16, which is exact. This entry packs attention and MLP in bfloat8_b, also exactly, including the fused q|k|v
and gate|up (checked on the first 2 layers):
"gemma-4-E4B-it-Tenstorrent-Tensix-BFP8": {"default": {"attention": "bfp8", "shared_mlp": "bfp8"}}
tt-forge (tt-xla). Not run with Gemma 4 yet. By default it keeps the weights in bf16, which is exact. Its
bfp_bf8 option re-quantizes them in PyTorch's [out, in] layout; bfp_bf4 quantizes them further.
Evaluation
Table with columns: Bits a weight, Size (GB), Mean KL to BF16, Top-1 (%), Answers right, Lost | Bits a weight | Size (GB) | Mean KL to BF16 | Top-1 (%) | Answers right | Lost |
|---|
| BF16 (google/gemma-4-E4B-it) | 16 | 15.04 | 0 | 100 | 36 of 40 | - |
| this model | 12.06 | 11.337 | 0.00169 | 98.12 | 36 of 40 | 0 |
| Q4_K_M (llama.cpp) | 5.66 | 5.335 | 0.19989 | 81.05 | 36 of 40 | 0 |
KL and top-1 cover 7,968 positions of English and Korean text (WikiText-2 test, Korean Wikipedia), in the chat
template. Answers are 33 short prompts with automatic checks, plus the 7 reasoning ones again with thinking on
(40 in all). "Lost" means right in BF16 but wrong or broken here; near ties are not counted. Q4_K_M was made
from the same checkpoint with llama.cpp v0.9.0 (cd959bae) and no importance matrix; its sha256 is
71766ed49554884dc89513c2787f67c360a1de92fb928078f6464c78b090b2e5. Our bar: mean KL at most 1.5 times
Q4_K_M's and at most one more lost answer. Passed: 0.01 times Q4_K_M's KL, 0 lost against 0.
These numbers measure closeness to BF16, not task quality. Speed is not reported here.
License
Apache 2.0, as the base model's card says; its repository has no LICENSE file, so LICENSE holds the standard
text. The changes are listed in NOTICE. pack_tensix.py and verify_grid.py are by Xylolabs under the same
license. Provided as is, without warranty. Not affiliated with or endorsed by Tenstorrent.
Citation
@misc{gemmateam2026gemma4, title={Gemma 4 Technical Report}, author={Gemma Team}, year={2026}, eprint={2607.02770}, archivePrefix={arXiv}, primaryClass={cs.CL}, url={https://arxiv.org/abs/2607.02770},}