IJ retweeted
This next DeepSeek 4.1 Flash update is a big one! Here are some of my favourite results so far.
- More context (Hopefully 1M x 4)
- Fewer mistakes and better tool calling
- Less thinking (12%) & 1/3 fewer "Confident mistakes"
- 20% better at non english language
- Another 2-5% faster
- >= MMLU (88-89%)
Full release will come in a few days with a full report on how I did it.