DeepSeek released the official build of V4 Flash today, named DeepSeek-V4-Flash-0731, and the interesting part is not that a new model dropped. It is what post training alone did to agent performance.
Same model, new job
The 0731 build keeps the exact architecture and size of the April preview: 284B total parameters, 13B active, 1M token context. DeepSeek only re ran post training, aimed squarely at agent workloads. The deltas:
| Benchmark | Preview | 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 without touching the base model is the kind of jump that usually needs a new generation. On several code and agent benchmarks the small Flash now outscores DeepSeek's own V4 Pro Preview, a 1.6T model activating nearly four times the parameters per token.
What that means for people building agents
Two shipping details matter as much as the scores. First, 0731 natively supports the Responses API format and is adapted for Codex, so agent tooling built on that interface runs against deepseek-v4-flash without a compatibility shim. Second, the pricing: $0.14 per million input tokens, $0.28 per million output, and $0.0028 on cache hits. Agent loops resend context constantly, which makes the cache hit price the number that decides your bill.
And the open part is not marketing: the weights landed on Hugging Face under MIT the same day, community GGUFs followed within hours, and the 155 GB 4 bit quant fits a 192 GB workstation or Mac Studio. A deep dive into what shipped, including the fine print (announced 2x peak hour pricing, the custom chat template, and where the model is unremarkable), is here: DeepSeek V4 Flash 0731 Explained.
Trying it without an API console
If you want to poke the model before wiring it into anything, the routes are: the official API (OpenAI compatible), OpenRouter (one model string), a flat price hosted plan such as LU Labs where V4 Flash and V4 Pro sit in one picker with monthly quota instead of per token billing, or fully local on your own hardware since the weights are open. The whole decision tree with working code samples is in the cloud guide, and the local path is covered in the local guide.
The strategic takeaway: specialization beat scale today. A 13B active model, trained hard on the agent loop, is outperforming a 49B active generalist on the workloads developers are building right now. Expect every lab to copy this recipe.
Top comments (0)