Unverified50% confidenceBenchmarkExact time
In 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%.
1
Sources
50%
Confidence
Long-term
Relevance
10/12/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedOn a three-sample math problem, DeepSeek V4.1 Flash scored 74 points, on par with Opus 5.0 and surpassing GLM's 67 points and V4 Pro's 62 points.75% similarUnverifiedMistral Small 4 scored only 0.699 on unit conversion inputs and produced 265 errors that differed from the correct label by exactly a power of ten.72% 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.71% similarUnverifiedThe parsing defect combined with one incorrect answer key caused gpt-oss's score across five core tasks to shift from 0.62 to 1.00, a shift as large as 0.38.70% similarUnverifiedAll incorrect distractor counting answers were without exception overestimates, and the token counts of incorrect answers were clearly higher than those of correct answers: for gpt-5.4-mini at high, the median token count of incorrect answers was 412 vs 91.5 for correct answers; for gpt-oss-120b, incorrect answers were 5,776 tokens vs 747 tokens for correct answers.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1012260API
curl https://kongchang.com/api/v1/knowledge/claims/1012260MCP
get_claim(id=1012260)