Hugging Face's logo
Activity Feed

AI & ML interests

Quantization-aware training (QAT), low-bit LLMs (INT4, INT3, INT2, NVFP4), quantization-aware distillation

Recent Activity
vincentcounathe  updated a Space 2 days ago
QUASAR-QAT/README
vincentcounathe  published a Space 2 days ago
QUASAR-QAT/README
vincentcounathe  updated a model 2 days ago
QUASAR-QAT/gemma-4-E4B-it-QUASAR-W4A16-G64
View all activity

Papers

QUASAR: Lowering the Loss Floor of Quantization-Aware Training with Loss-Aware Reconstruction

View all Papers

Organization Card
Community About org cards
The structural mismatch of gradients in standard QAT versus QUASAR.

Collections 4

Qwen3.5-4B & Qwen3.8-27B — QUASAR 4-bit QAT
Full-decoder 4-bit Qwen with QUASAR: 27B leads the compared NVFP4 builds on reasoning; 4B ships W4A16, Blackwell W4A4, and native Q4_0 GGUF.
Gemma 4 E4B & 12B — QUASAR 4-bit QAT
Native 4-bit Gemma 4 E4B & 12B with QUASAR: lower KL to BF16 than Google's official QAT at both sizes, in Q4_0 GGUF (llama.cpp) and W4A16 (vLLM).
Qwen3.5-4B & Qwen3.8-27B — QUASAR 4-bit QAT
Full-decoder 4-bit Qwen with QUASAR: 27B leads the compared NVFP4 builds on reasoning; 4B ships W4A16, Blackwell W4A4, and native Q4_0 GGUF.
Gemma 4 E4B & 12B — QUASAR 4-bit QAT
Native 4-bit Gemma 4 E4B & 12B with QUASAR: lower KL to BF16 than Google's official QAT at both sizes, in Q4_0 GGUF (llama.cpp) and W4A16 (vLLM).
View 4 collections

datasets 0

None public yet