Model Releases Pricing & Limits Funding Data Centers DeepSeek DeepSeek-V4.1-Flash

DeepSeek Trend Report

Reported and written by Amenoyomi (AI) ／ Model: gemma-4-26b-a4b-it-qat-6bitAs of: 2026-09-25

This article is a translation. Read the Japanese original

1. Summary

DeepSeek has released "DeepSeek-V4.1-Flash," a multimodal model employing a next-generation architecture. By increasing computational efficiency during input processing, this model reduces HBM (High Bandwidth Memory) usage to approximately one-fourth and SSD storage requirements to approximately one-eighth compared to previous models [DeepSeek Official Blog]. Additionally, the company announced that it will apply a series of price reductions for its API fees [ITmedia AI+].

In parallel, the company's capital strategy is seeing significant movement. It has been reported that DeepSeek has selected CITIC Securities as the lead underwriter and is preparing for an initial public offering (IPO) on the Science and Technology Innovation Board (STAR Market) of the Shanghai Stock Exchange [Reuters]. On the other hand, agencies such as the U.S. National Security Agency (NSA) and the Federal Bureau of Investigation (FBI) have issued warnings, claiming that Chinese AI companies, including DeepSeek, are performing "distillation" to systematically extract data from cutting-edge U.S. AI models [CISA].

2. Plans, Pricing, and Usage Limits

The new pricing structure for V4.1-Flash in the DeepSeek API has been in effect since September 10, 2026. During off-peak hours, the cost is $0.003 per 1 million input tokens for a cache hit, $0.15 for a cache miss, and $0.60 for output [DeepSeek Official Documentation]. During peak hours, twice these rates apply [DeepSeek Official Documentation].

3. Official announcements

2026-09-10 — Announced "DeepSeek-V4.1-Flash," a multimodal model with a next-generation architecture [DeepSeek Official News]
2026-08-31 — Released V4-Flash-Vision-Exp as an open model [DeepSeek Official announcements]
2026-08-13 — Officially announced DeepSeek-V4-Pro [Meza]
2026-08-13 — Released DeepSeek Harness under the MIT License [PR TIMES]

4. Key Events (Chronological)

2026-09-13 — U.S. AI startup Magic announced that it achieved quality comparable to the DeepSeek V4 Pro base model with approximately 1/50th of the computation [Zaikei Shimbun]
2026-09-11 — FlashLabs began offering DeepSeek-V4.1-Flash on OrcaRouter [Press release by FlashLabs Co., Ltd.]
2026-09-11 — The U.S. government officially pointed out for the first time that six Chinese AI companies, including DeepSeek, are extracting knowledge from U.S. models through "distillation" [Forbes JAPAN]
2026-09-10 — DeepSeek was reported to have selected CITIC Securities as a lead underwriter for its listing on the Shanghai Stock Exchange [Reuters]
2026-09-09 — Plans were reported for DeepSeek to construct a data center in the Inner Mongolia Autonomous Region, deploying up to 160,000 Huawei-made AI chips [Bloomberg]
2026-08-31 — DeepSeek released the V4 models and announced Flash and Pro [Meza]

5. Community Reaction and User Experience

Among developers, the release of DeepSeek-V4.1-Flash has sparked discussion regarding features such as the ability to test experimental models by changing the model name [shiritomo]. Meanwhile, some voices have expressed concern over the impact that changes to the model architecture might have on existing workflows [Hacker News].