Unverified50% confidenceBenchmarkExact time
In 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.
1
Sources
50%
Confidence
Long-term
Relevance
10/12/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIn the task of completing the last Token of an FP8 quantization function, after DeepSeek-V4-Flash-Base adds equal-sign decorations before the code, when the number of equal signs modulo 4 is 0 or 1 the error rate is as high as 71.3%, and when the remainder is 2 or 3 the accuracy returns to 91.5%.71% similarUnverifiedOn the Aider Polyglot coding test, DeepSeek at full precision scored 71.6%, Unsloth's 4-bit dynamic version scored 69.7%, the 2-bit version scored 65.8%, while a competing quantization solution of the same 2-bit depth and file size scored only 56.6%.69% similarUnverifiedIn the GPQA-Diamond xhigh test, the base model Qwen3.8-27B achieves 88.4% accuracy consuming 6,642 median tokens, while Swift maintains 88.3% accuracy with tokens reduced to 2,77166% similarUnverifiedIn DeepSeek's official comparison, when running with the same checkpoint at best effort, its minimal framework scored 72.6 while the open-source code scored 65.5.66% similarUnverifiedAccording to Artificial Analysis leaderboard data, DeepSeek V4 Pro's score increased from 52 to 53, a difference of only 1 point.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1012262API
curl https://kongchang.com/api/v1/knowledge/claims/1012262MCP
get_claim(id=1012262)