I asked DeepSeek V4.1 Flash to write code.
By John Aspinall ✱ · September 14, 2026 · Curated by George's Blog
I asked DeepSeek V4.1 Flash to write code.
It refused.
Told me my architecture was flawed.
⤷ Then it fixed it.
Not kidding here.
On September 10, DeepSeek dropped V4.1 Flash. A smaller model. Beats their flagship V4 Pro so decisively they're killing it .
I spent 72 hours testing.
Three moments stopped me cold.
1. I asked for a data script. It gave me code + a warning:
"Your field names are ambiguous. Three interpretations below. Confirm first."
2. I gave it a prompt with a logical trap.
It didn't fall for it. Called out the flaw directly.
3. It refused a task.
Reason: "This will leak data. I don't recommend it."
Old AI: "You say it, I do it."
New AI: "I understand. I judge. I refuse."
552B parameters.
8B active reading.
16B generating.
Native vision. Built in. Not bolted on.
So ask yourself:
Do you want your AI obedient or intelligent?
If it refuses your instruction, is that "broken" or "trustworthy"?