Table with columns: Tensors, Format| Tensors | Format |
|---|
| attention q, k, v, o and MLP gate, up, down (36 layers) | bfloat4_b, 4.5 bits a weight |
| LM head (tied to the embedding), embedding, norms and every other tensor | bf16, as in the base model |
5.61 bits a weight on average, 2.82 GB packed. quant4.json lists each projection's format. Quantized with
GPTQ and a per-block exponent search, calibrated on 65,536 tokens of WikiText-2 and Korean Wikipedia.
Quickstart
from transformers import AutoModelForCausalLM, AutoTokenizer repo = "Xylolabs/Qwen3-4B-Tenstorrent-Tensix-BFP4"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 = "Qwen3-4B-Tenstorrent-Tensix-BFP4" # hf download Xylolabs/Qwen3-4B-Tenstorrent-Tensix-BFP4 --local-dir Qwen3-4B-Tenstorrent-Tensix-BFP4name = "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-Tenstorrent-Tensix-BFP4 checks every matrix this way (--device also round-trips it
through device DRAM). python verify_grid.py Qwen3-4B-Tenstorrent-Tensix-BFP4 checks the grid with torch alone. The same
values also pack exactly into ttnn.bfloat8_b.
tt-transformers (tt-metal v0.81.1). Pass --use_hf_rope; without it the fused QKV is re-quantized. The
demo packs the tied head in bfloat8_b, so the 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-Tenstorrent-Tensix-BFP4export TT_CACHE_PATH=$PWD/model_cache/Qwen3-4B-Tenstorrent-Tensix-BFP4pytest models/tt_transformers/demo/simple_text_demo.py -k "performance and batch-1" --use_hf_rope \ --optimizations 'precision_cfg = {ff1_3: bfp4, ff2: bfp4, wqkv: bfp4, wo: bfp4, 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). It runs the weights in bf16 by default, which is exact. Its bfp_bf4 option
re-quantizes them in PyTorch's [out, in] layout, so use tt-transformers for 4.5-bit weights. Not yet run end
to end.
pip install pjrt-plugin-tt==1.4.0 --extra-index-url https://pypi.eng.aws.tenstorrent.com/tt-forge-installpython - <<'PY'import torch, torch_xla, torch_xla.runtime as xrfrom transformers import AutoModelForCausalLM, AutoTokenizerxr.set_device_type("TT")device = torch_xla.device()tok = AutoTokenizer.from_pretrained("Xylolabs/Qwen3-4B-Tenstorrent-Tensix-BFP4")model = AutoModelForCausalLM.from_pretrained("Xylolabs/Qwen3-4B-Tenstorrent-Tensix-BFP4", torch_dtype=torch.bfloat16).eval().to(device)compiled = torch.compile(model, backend="tt")ids = tok("The capital of France is", return_tensors="pt").input_ids.to(device)with torch.no_grad(): logits = compiled(input_ids=ids).logits[0, -1].cpu()print(tok.decode(logits.argmax()))PY
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 | 5.61 | 2.822 | 0.0711 | 89.30 | 36 of 40 | 1 |
| 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: 1.08 times Q4_K_M's KL, 1 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 (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},}