needle-1M-bench + Qwen3 quantizations
Collection
Long-context faithfulness benchmark + audit-friendly Qwen3 quantized releases. Outputs ship; inputs are auditable. • 23 items • Updated
Qwen3-14B-HQQ-INT4
INT4 quantization of Qwen/Qwen3-14B. Calibration-free companion to drawais/Qwen3-14B-AWQ-INT4.
Footprint
| Source params | 14B |
| Quantized weights | ~9.5 GB on disk |
| Inference VRAM (incl. KV cache @ 32K context) | ~16 GB |
Recommended use
Best at native 40K context. This release is intended for chat / interactive workloads at the base model's native context length.
For long-context (>40K) use the AWQ companion — drawais/Qwen3-14B-AWQ-INT4, which scored 90% overall on needle-1M-bench-mvp 50K with YaRN extension via vLLM.
Bench
The companion AWQ release scored 90.0% overall (paper-anchored 80%, synthetic 100%) on drawais/needle-1M-bench-mvp 50K with YaRN-extended context via vLLM. A direct HQQ score at native ≤40K context is queued for a future release.
Quick start
from transformers import AutoTokenizer
from hqq.models.hf.base import AutoHQQHFModel
tok = AutoTokenizer.from_pretrained("drawais/Qwen3-14B-HQQ-INT4")
model = AutoHQQHFModel.from_quantized("drawais/Qwen3-14B-HQQ-INT4", device="cuda")
License
Apache 2.0 (inherits from base model).
- Downloads last month
- 43