What is DeepSeek V4 Flash 0731?

The official release build of DeepSeek V4 Flash, published July 31, 2026. Same architecture and size as the preview (284B total parameters, 13B active, 1M context), but re post trained with a heavy focus on agent workloads. It replaces the preview on the API under the same model name, deepseek-v4-flash, and the weights are on Hugging Face under MIT.

How good is DeepSeek V4 Flash 0731 at agent tasks?

The jump over the preview is dramatic: Terminal Bench 2.1 went from 61.8 to 82.7 and DeepSWE from 7.3 to 54.4. On several code and agent benchmarks it now beats DeepSeek's own larger V4 Pro Preview despite activating a fraction of the parameters.

Is DeepSeek V4 Flash 0731 open source?

The weights are open under the MIT license on Hugging Face and ModelScope, published the same day as the API release. GGUF quantizations from the community appeared within hours.

Did the DeepSeek app get the new model too?

Not yet. DeepSeek states that only the API was upgraded to 0731; the app and web versions still run the previous models. The V4 Pro API is also unchanged.

DeepSeek V4 Flash 0731 Explained: Why Today's Release Is a Big Deal

DeepSeek released the official build of V4 Flash today, July 31, 2026, under the name DeepSeek-V4-Flash-0731, and the hype is loud even by model launch standards. The short version: same 284B parameter MoE as the preview, re post trained into one of the strongest agent models available anywhere, open MIT weights on day one, and prices that undercut everything in its class. Here is what changed and why it matters.

The Numbers Driving the Hype

The preview of V4 Flash, released in April, was a solid mid tier model. The 0731 build is something else. DeepSeek only re ran post training, same architecture, same size, and the agent scores moved like this:

Benchmark	V4 Flash Preview	V4 Flash 0731
Terminal Bench 2.1	61.8	82.7
DeepSWE	7.3	54.4
NL2Repo	n/a	54.2

DeepSWE going from 7.3 to 54.4 through post training alone is the kind of jump that usually needs a new base model. And on several code and agent benchmarks, this 13B active model now beats DeepSeek's own V4 Pro Preview, a 1.6T parameter model activating 49B per token. Reasoning depth still favors the big one (see Flash vs Pro), but for the agent loop workloads everyone is building, small and specialized just beat big and general.

What Shipped Today

API: the model name stays deepseek-v4-flash; requests now hit the 0731 build. The app and web versions are unchanged, and so is the V4 Pro API.
Responses API and Codex: 0731 natively speaks OpenAI's Responses format and is adapted for Codex, so agent tooling built on that interface works without a shim. That is new for DeepSeek and clearly aimed at the coding agent market.
Open weights: published same day on Hugging Face and ModelScope under MIT, with community GGUFs live within hours. The preview took this path too, and its weights became some of the most downloaded of 2026.
Reasoning modes: thinking is the default, with a non thinking mode per request, and the architecture (hybrid compressed sparse attention) keeps the 1M token context from the preview.

Why a 13B Active Model Matters More Than a 1.6T One

Benchmarks aside, the shape of V4 Flash is the strategic part. 284B total with 13B active means: cheap to serve ($0.14 per million input, $0.28 out, and $0.0028 on cache hits), fast enough for interactive agents, and small enough that a serious workstation or a 192 GB Mac Studio can hold the 4 bit quant. The frontier agent model of the day is one you can download and own. That combination did not exist before today.

How to Try It Right Now

On your own hardware: the full walkthrough, from the 5 GB distill to the 155 GB GGUF, is in How to Run DeepSeek V4 Flash Locally. The V4 Flash family is in the one click catalog of our free local AI studio.
In the cloud: LU Labs hosted plans, the official API with code samples, and OpenRouter are covered in the cloud guide.
Uncensored: the abliterated build of V4 Flash is the most downloaded uncensored model of 2026; details in this post.

The Catch List

For balance, the things to keep an eye on: peak hour API pricing at 2x is announced (Beijing business hours, start date pending); the app and web products still serve older models, so casual users are not seeing 0731 yet; HumanEval style single shot code scores are unremarkable for the class (69.5), it is the multi step agent work where this build shines; and the chat template is custom (DeepSeek ships encoding scripts instead of a standard Jinja template), which tripped up some third party integrations on day one.