The Kohnex blog
Ideas worth
thinking about.
On reasoning, evidence, and a more considered kind of intelligence.
Same model. A different dimension.
DeepSeek V4 Flash went from 42.7% to 65.2% on 100 custom expert tasks. Here is the full picture.
Read the story
Proof over promises. Always.
Full breakdowns. The unflattering numbers, too. Our evaluation policy, in writing.
Read the story
Thinking bigger isn’t about size.
What a compact model reveals about structure, scale, and the future of expert reasoning.
Read the story