Skip to content
China AI DispatchChina AI news, every weekday

Models / DeepSeek-V4-Flash-0731

DeepSeek

DeepSeek-V4-Flash-0731

Specs as of

DeepSeek’s V4-Flash-0731, released 31 Jul 2026: 284B total and 13B active parameters, a 1M-token context, open weights, and what its benchmarks show.

Specifications

Lab
DeepSeek
Released
31 Jul 2026
Parameters
284B total, 13B active (MoE)
Context window
1M tokens
License
Open-weight
Pricing
About $0.14 per million input tokens and $0.28 per million output, roughly one fifty-fourth the price of Claude Opus 4.8. On 6 Aug 2026 DeepSeek warned of a significant price increase, a first for the company that started China’s model price war, with co-founder Jun Song saying that even after a two-to-tenfold rise it would still sit below most Western models

The benchmark claim

On DeepSeek’s own nine-benchmark agent-and-coding suite it lands close to Claude Opus 4.8, but Opus 4.8 still leads every one of the nine (for example 85.0 to 82.7 on one terminal-agent test and 71.7 to 59.6 on a harder data-science set); a hands-on Bilibili test found it usable for everyday coding but faltering on complex 3D work, as expected from 13B active parameterslab-reported

A figure tagged lab-reported is the lab's own claim. A figure taggedindependent comes from a third-party leaderboard. We never present a vendor's number as neutral fact.

Sources

DeepSeek benchmark suite; officechai; Bloomberg; SCMP; TechNode; Bilibili (AI超元域)

Covered in the daily: read the issue →

Every figure above traces to the named source on this page, and where it comes from our reporting, to the dated issue that carried it. Specs the lab did not disclose read "n/d" rather than carrying an estimate.

Subscribe

The stories English coverage misses, every morning

I scan 100+ Chinese-language sources every day and write up the China AI stories English coverage misses. It goes out every weekday. Free.

Free. One email each weekday morning. Unsubscribe anytime.