Unverified50% confidenceBenchmarkExact time
When DINOv3 ViT-H+/16 is pruned to 50% sparsity, removing 427 million parameters, ImageNet-1k linear probe accuracy drops by only 0.8%, without weight correction or retraining.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/15/2026
First Seen
Valid until: 1/13/2027
Sources
Related Entities
Related Claims
UnverifiedNo Decision 3.0 model was exposed to video data during training, yet on all 19,140 questions in the Perception Test validation set, the 27B version achieved a zero-shot score of 77.4 and the 4B version scored 70.3.69% similarUnverifiedIn a needle-in-a-haystack experiment with a 128K Token context and about 16,000 key-value pairs, changing the position of just one key-value pair produced a maximum accuracy gap of 40.2% for DeepSeek-V4-Flash-Base, 34.8% for DeepSeek-V4-Pro-Base, and still 19.1% and 14.8% respectively after post-training optimization.68% similarUnverifiedIn the Qwen3-4B comparative experiment, updating the model directly with unfiltered skills without using SGUID filtering decreased performance by 0.3 percentage points on HMMT25.67% similarUnverified一个rank=4的LoRA文件对于SD1.5模型约只有3-6MB，LoRA设计使训练参数量减少95%以上67% similarVerifiedGGUF量化格式（如Q4_K_M）通过将权重精度从32位浮点压缩到4位整数，通常只带来约1-5%的性能损失66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1037994API
curl https://kongchang.com/api/v1/knowledge/claims/1037994MCP
get_claim(id=1037994)