Unverified50% confidenceFactExact time
The training cost of the Qwen4 architecture is only one-ninth that of the previous generation.
1
Sources
50%
Confidence
Long-term
Relevance
10/29/2026
First Seen
Sources
Related Entities
Related Claims
VerifiedWhen fine-tuning with LoRA, the actual number of parameters that need to be trained is only 0.1%-1% of the original.63% similarUnverifiedQwen 3.8 is slower than 3.6, with prefill and inference speeds dropping by approximately 40% and 20% respectively.58% similarUnverifiedAcross the Qwen3 model scale series, APSD reaches performance comparable to SDPO with an average of 2.6 times fewer training steps.58% similarUnverifiedThe inference cost of the new generation of code models is only one-quarter that of the previous generation.57% similarUnverifiedDiffusionGemma的监督微调步骤成本不到原模型训练总开销的10%57% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1112068API
curl https://kongchang.com/api/v1/knowledge/claims/1112068MCP
get_claim(id=1112068)