Unverified50% confidenceTradeoffExact time
The same Laya weights scored 55 on llama.cpp but only 41 on PyTorch and MLX; changing the JSON state to labeled text raised the Laya judge group score from 1/10 to 4/10.
1
Sources
50%
Confidence
Long-term
Relevance
10/11/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIn the MMLU Pro test, 8-bit narrowly beat OQ4E with 60.6% versus 59.9%.56% similarUnverifiedLaya achieves only 0.45 accuracy on the MASSIVE dataset covering 13 languages and only 0.425 on Banking77, while JAF reports 0.87 on the same tasks.56% similarUnverifiedLlama 3.3 70B's kinship term selection accuracy was 77.92%, while its generation accuracy was only 24.24%.55% similarUnverifiedThe Laya decision model responds up to 4.1 times faster for short requests, with a 3-question guardrail check on a B200 dropping from 11.6 ms to 2.8 ms.54% similarUnverifiedThe Qwen4 preview version scored only 35.9 on the HLE reasoning leaderboard, losing to a competitor's enhanced version that scored 40.54% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/998812API
curl https://kongchang.com/api/v1/knowledge/claims/998812MCP
get_claim(id=998812)