DeepSeek updated its V4.1 Flash model yesterday. One obvious change is that it shortened the output length, making answers faster and cheaper.
I previously wrote two pieces on this issue, and the most common criticism I faced was that I treated chain-of-thought as output as material for criticism, and I got severely criticized for it.
Before this, using it to assist with programming often meant being tormented by screens full of redundant text. If I wanted a simple configuration, it would insist on laying out a preamble in the order of what it checked, how it analyzed it, and how it verified it, with the main content scattered across every corner of the answer.
This kind of repetitive boilerplate is not because the model’s capability is lacking, but because the evaluation criteria during training mistook step-by-step self-justification for high-quality answers. To prove it did the work, the model tends to lay out the entire reasoning chain and then casually pile on a large number of just-in-case disclaimer assumptions.
In model evaluation systems, answers that explain all details usually score high. However, in real engineering scenarios, the more fully the reasoning steps are written in an answer, the higher the reading cost, and actual delivery efficiency instead falls off a cliff.
This leads to a mismatch: when actually doing the work, I only care about the final result, while the model is justifying its logic.
The official team evidently also recognized this problem. The new version uses reinforcement learning to correct this metric distortion and proactively cut out a large amount of nonsense.
The new version’s score on Terminal-Bench rose directly from 31.2 to 42.3. Terminal-Bench has high requirements for model output accuracy; the more output and the more redundant it is, the more likely it is to cause command distortion, and the lower the score.
You could say this update is entirely aimed at the problem I raised. This is genuine, real experience; people without heavy usage experience cannot appreciate it.