Unverified50% confidenceFactExact time
The QLoRA reward margin differs by up to 3.4× between nf4 quantized training and unquantized fp16 measurement, and the gap depends on the measurement axis.
1
Sources
50%
Confidence
Long-term
Relevance
11/5/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIn FP8 mixed-precision training, V3 uses FP8 computation for most matrix multiplications and retains higher precision for precision-sensitive steps, theoretically doubling compute throughput and halving memory usage compared with BF16.71% similarUnverifiedGradient checkpointing and mixed-precision training (FP16/BF16) are common memory-reduction measures.70% similarUnverifiedUsing QLoRA on A100 and RTX PRO 6000, training speed for Qwen3.5-35B-A3B can be improved by up to 4.1×, and for Qwen3-30B-A3B by up to 3.3×.68% similarUnverified3,189 successful rollouts were collected from Qwen3.8-27B for supervised fine-tuning (imitation learning), and performance stagnated at 47.5%, below the results of two rounds of RL training.68% similarUnverifiedThe performance gap between a carefully quantized 4-bit model and the original FP16 model is typically within 1–3 percentage points.68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1138290API
curl https://kongchang.com/api/v1/knowledge/claims/1138290MCP
get_claim(id=1138290)