Unverified50% confidenceFactExact time
NVIDIA's FP4 version of Qwen3.8 was trained using quantization-aware distillation (QAD), with a training budget of 3 epochs totaling 500M tokens, comprising 2,500 backbone updates and 1,500 joint fine-tuning updates.
1
Sources
50%
Confidence
Long-term
Relevance
10/17/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedNodus 在 2026 年 10 月 11 日的试点中，使用 Qwen2.5-0.5B 模型在单张 NVIDIA L4 上运行相同配方，监督微调费用比 Modal 低 67%，持续预训练费用低 42.5%；计入一次失败尝试后微调节省降至 24.7%。70% similarUnverifiedQwen4, built on a next-generation architecture, has already entered training.70% similarUnverifiedOn seven benchmarks covering mathematics, code generation and dialogue, ReaDy using Qwen3.8-27B achieved an average speedup of 3.16x over autoregressive decoding on a single NVIDIA B200 GPU.69% similarUnverified混合精度训练由 NVIDIA 在 2018 年提出，同时使用 FP32 和 FP16 两种精度69% similarUnverifiedQwen2.5系列在后训练阶段引入了大量工具调用对齐数据，并通过RLHF和DPO强化了结构化输出的可靠性68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1049818API
curl https://kongchang.com/api/v1/knowledge/claims/1049818MCP
get_claim(id=1049818)