NariQ trellis quantization matches Qwen3.8-27B BF16 on long tasks
Researchers have proposed NariQ, a trellis-coded post-training quantization method designed to address the performance drop of low-bit models on long-horizon agentic tasks. Calibrated on tokens generated by the full-precision model during agent sessions, NariQ maintains performance parity with the BF16 version of Qwen3.8-27B at 3.0 and 2.5 bits per weight across multiple agentic and general benchmarks. The team has released the corresponding GGUF checkpoints and a llama.cpp fork for inference. Open article →