Unverified50% confidenceBenchmarkExact time
After three rounds of fine-tuning, a 322M-parameter model's allow rate on everyday commands rose from 25/50 for the base model to 44/50 in the third round, and its attack interception rate rose from 32/43 to 39/43.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
11/3/2026
First Seen
Valid until: 2/1/2027
Sources
Related Entities
Related Claims
UnverifiedIn the evaluation, enabling the full context window (256K) raised the small model's task success rate from 26% to 53%.65% similarUnverifiedV4-Pro的总参数与活跃参数比例约为32:1的稀疏比64% similarUnverifiedIn tests on 2,456 held-out prompts, increasing the sampling steps from 20 to 50 improved FID only from 27.0 to 26.6 (a gain of 0.4), while removing the shift caused FID to rise from 27.0 to 27.3 and FD-DINOv2 to rise from 218 to 228.63% similarUnverifiedThe H3 model is trained with resolutions processed as multiples of 32; keeping Multiple at 32 during inference avoids artifacts or performance degradation caused by size mismatches.63% similarUnverifiedA model with approximately 27B parameters achieved a score of 52 running on a laptop, demonstrating that high-quality models are penetrating edge and lightweight devices.63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1130039API
curl https://kongchang.com/api/v1/knowledge/claims/1130039MCP
get_claim(id=1130039)