If I Don’t Ask, It Doesn’t Act

I recently switched to a new model for work: deepseek-v4-flash-0731.

After using it for a few days, I’ve figured out its temperament: if I don’t ask, it doesn’t act.

Tell it to fix a bug, and it fixes only that bug. Three lines away there might be an obviously explosive bit of code, but if I don’t call it out, it won’t touch it. Once it’s done, it won’t ask a follow-up either—things like “Want me to add a test while I’m here?” are simply not in its vocabulary.

What’s the Story Behind 0731?

First, a bit of context.

The 0731 in the name refers to July 31, the day the official version of DeepSeek V4 went live. Flash is the fast tier in this generation. A preview version came out in late April, and 0731 is basically the graduation to a stable release.

The architecture is called MoE, or Mixture of Experts. It has 284 billion total parameters, but only wakes up 13 billion for any given task—like a company with hundreds of experts on staff, where you ask a question and the front desk calls in only the few most relevant people. That’s why it’s cheap, fast, and can still handle a 1-million-token context window, enough to stuff in an entire long novel.

The official positioning is also very straightforward: this is a workhorse model. If you really need to chew through the hard stuff, like complex planning, they tell you to go to its big brother, V4-Pro.

It Only Does What I Explicitly Name

What does “if I don’t ask, it doesn’t act” feel like? Here are two examples.

I asked it to add a feature to a script. It added it, and the feature worked. But the function next to it had a terrible name, and the error handling was completely naked—some other models would have already rewritten the whole thing for me in passing. This one didn’t touch it.

Another time, I said, “Take a look at this config and see if there’s anything wrong.” So it looked only at the config. The script referenced inside the config had problems? Not its job.

It’s not that it never does anything extra, but most of the time, it only does what I explicitly name. At first I thought it was lazy. Later I realized: it isn’t lazy, it’s just extremely obedient. The boundary of the task is exactly the sentence I say out loud—not a single word more.

Whether That’s a Flaw or a Feature Depends on You

If you’re used to throwing out a vague “help me fix this” and then walking away, you’re going to have a bad time. Its understanding of “fix this” is the most literal possible version of “fix this.”

But my habit with prompts has always been to spell everything out at once: what to change, what not to change, and what counts as done. Used that way, its passivity actually gives me peace of mind—it won’t quietly refactor unrelated files while I’m not paying attention. Every step it takes stays inside the circle I drew.

I’ve written before: AI is the executor; I’m the commander. This model’s most remarkable trait is exactly that: if you don’t command it, it really won’t move.

Cheap Means Actually Cheap

There’s another very practical upside: it’s fast, and it’s cheap.

Didn’t give complete instructions? Run it again—no pain. Throw a few hundred instructions at it in a day, and the bill barely registers. Compared with models where you toss in one sentence, they improvise freely for three minutes, and the bill improvises right along with them, this rhythm of “I say one thing, it does one step” suits me much better.

Finally

Some people might ask: doesn’t it bother you that it isn’t proactive enough?

Not at all. Proactivity is something I already have. What I need is someone that executes what I say cheaply and accurately.

It won’t think for me. Thinking was always my job to begin with.