TQ: calibration-free 4-bit quantization open-sourced, hits 92.4% top-1 on Qwen 27B
textclf · reddit · 2026-09-28
A developer open-sourced TQ, a calibration-free quantization method, together with its implementation 'Quant Factory':
- Key claim: needs no calibration data, so models can be quantized the moment they ship, and quantized weights generalize better while staying close to calibration-based methods
- Numbers: 4-bit TQ on Qwen 3.8 27B achieves mean KLD of 0.0282 and 92.4% top-1
- Method: grounded in information-theoretic lossy source coding, claimed optimal at the rate-distortion limit
- Practicality: 4-bit only for now (2-bit/3-bit planned); quantized models published on Hugging Face, with a Docker image that serves them via vLLM using --quantization tq
More from Infra
- Gemini 3.5 Flash now runs on air-gapped Google Distributed Cloud, offline up to 12 months — shashib · 2026-11-16
- Gemma.c: full Gemma 1/2/3 multimodal inference in a single C file, ~72 tok/s on CPU at INT8 — This-Purchase-3325 · 2026-11-16
- Sarvam AI Launches Token Factory: India-Hosted Inference for Open Models — itsOmSarraf_ · 2026-11-16
- BIS Issues Interim Final Rule Addressing Remote Access to Advanced AI Chips — apples_jimmy · 2026-11-16
- Jailbroken PS4 runs a 9B open-source model at up to 43 tok/s by talking directly to its 2013-era GPU — AIFlow_ML · 2026-11-16
- Mobile AI Factory: agents auto-build on-device SLMs with 19x less GPU time — pmttyji · 2026-11-16