Table with columns: Tensors, Format| Tensors | Format |
|---|
| attention q, k, v, o and MLP gate, up, down (36 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 |
9.23 bits a weight on average, 4.64 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.
Quickstart
from transformers import AutoModelForCausalLM, AutoTokenizer repo = "Xylolabs/Qwen3-4B-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
Tested with ttnn 0.81.1 on ttsim, Tenstorrent's Blackhole simulator; not on hardware. 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 = "Qwen3-4B-Tensix-BFP8" # hf download Xylolabs/Qwen3-4B-Tensix-BFP8 --local-dir Qwen3-4B-Tensix-BFP8name = "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 Qwen3-4B-Tensix-BFP8 checks every matrix this way (--device also round-trips it
through device DRAM). python verify_grid.py Qwen3-4B-Tensix-BFP8 checks the grid with torch alone.
tt-transformers (tt-metal v0.81.1). Pass --use_hf_rope; without it the demo permutes the q and k rows
and re-quantizes the fused QKV. With these settings every projection of the first 2 layers packed exactly, and
a 2-layer prefill matched transformers (PCC 0.9990). The demo packs the tied head in bfloat8_b, so this
checkpoint's bf16 head is not exact there.
git clone --branch v0.81.1 https://github.com/tenstorrent/tt-metal.git && cd tt-metal# install tt-metal v0.81.1 as its INSTALLING.md describes, thenpip install -r tt_metal/python_env/requirements-dev.txtexport HF_MODEL=Xylolabs/Qwen3-4B-Tensix-BFP8export TT_CACHE_PATH=$PWD/model_cache/Qwen3-4B-Tensix-BFP8pytest models/tt_transformers/demo/simple_text_demo.py -k "performance and batch-1" --use_hf_rope \ --optimizations 'precision_cfg = {ff1_3: bfp8, ff2: bfp8, wqkv: bfp8, wo: bfp8, kv_cache: bfp8, activation: mixed}, fidelity_cfg = {li_ff1_3: hifi2, li_ff2: lofi, li_qkv_decode: hifi2, li_o_decode: hifi2, sdpa_decode: hifi2na, li_qkv_prefill: hifi2, li_o_prefill: hifi2fp16, sdpa_prefill: hifi4}'
tt-forge (tt-xla). Not run with this checkpoint yet. By default it keeps the weights in bf16, which is
exact. Its bfp_bf8 and bfp_bf4 options re-quantize them in PyTorch's [out, in] layout.
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 (Qwen/Qwen3-4B) | 16 | 8.04 | 0 | 100 | 36 of 40 | - |
| this model | 9.23 | 4.639 | 0.00120 | 98.63 | 36 of 40 | 0 |
| Q4_K_M (llama.cpp) | 4.95 | 2.497 | 0.06581 | 89.61 | 38 of 40 | 0 |
KL and top-1 cover 16,240 positions of English and Korean text (WikiText-2 test, Korean Wikipedia), plain and
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 f40a41777de3c1aaf31470ca19e004cc6579dfb57043b4f0e58a3701393d35b8. Our bar: mean KL at most 1.5
times Q4_K_M's and at most one more lost answer. Passed: 0.02 times Q4_K_M's KL, 0 lost against 0.
These numbers measure closeness to BF16, not task quality; no device speed was measured. For the model itself,
see the base card.
License
Apache 2.0, as the base model (LICENSE). 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{qwen3technicalreport, title={Qwen3 Technical Report}, author={Qwen Team}, year={2025}, eprint={2505.09388}, archivePrefix={arXiv}, primaryClass={cs.CL}, url={https://arxiv.org/abs/2505.09388},}