With today's V4 Flash 0731 release, DeepSeek's V4 family now has two very different members, and picking between them is a real decision. Short version: Flash for agents and code, Pro for the hardest reasoning, Flash by forfeit if you want to run it yourself.
The spec sheet
| V4 Flash (0731) | V4 Pro (Preview) | |
|---|---|---|
| Total params | 284B | 1.6T |
| Active per token | 13B | 49B |
| Context | 1M in / 384K out | 1M in / 384K out |
| API price per M | $0.14 in / $0.28 out | $0.435 in / $0.87 out |
| Responses API / Codex | Yes, native | No |
| Open weights | MIT | MIT |
| 4 bit GGUF | ~155 GB | ~900 GB |
The plot twist
Until this morning the story was: Pro is the smart one, Flash is the cheap one. The 0731 re post training broke that. Flash now posts 82.7 on Terminal Bench 2.1 and 54.4 on DeepSWE, and beats the V4 Pro Preview on several code and agent benchmarks while activating a quarter of the parameters. Pro keeps the lead where breadth and depth dominate: hard math, long proofs, dense multi document analysis over the full context window.
Cost wise the sticker gap says 3.1x on output, but agent reality widens it: agent loops burn output tokens (Flash's cheap side) and resend context (near free on cache hits for both). A day of heavy agent work that costs single digit dollars on Flash lands in the tens on Pro.
The local angle
This one is not close. Flash fits in about 155 GB at 4 bit and runs on a 192 GB Mac Studio or a big RAM workstation at interactive speeds, because only 13B parameters fire per token. Pro at 4 bit approaches 900 GB, which is multi GPU server territory and stays rented for almost everyone. If owning your model matters, the family decision is already made.
Picks by workload
- Coding agents, terminal work, repo tasks: Flash 0731. It benchmarks ahead and speaks the Responses API that Codex style tools expect.
- Everyday chat and writing: Flash. The quality gap does not justify 3x.
- Hard math and research grade reasoning: Pro.
- Full 1M context document analysis: Pro, hosted.
- Your own hardware: Flash, by forfeit.
The cheapest way to run the comparison yourself is a flat plan where both sit in one model picker, for example LU Labs Hosted Pro, so switching per task is free and you notice which model you stop switching away from. The full head to head with sources is here: DeepSeek V4 Flash vs V4 Pro, the Pro specific options (including the self hosting math) here: How to Run DeepSeek V4 Pro, and the local setup for Flash here: How to Run DeepSeek V4 Flash Locally.
Top comments (0)