DeepSeek V4.1 Flash Enters a Two-Day API Beta: New Architecture, Native Multimodal, and an Expiry Date in the Model Name

By the Deep Seek AI editorial desk · September 8, 2026 · 6 min read

DeepSeek did what it usually does before an announcement: it dropped the model into the live API. Around 15:00 Beijing time on September 8, 2026, the team told its official community groups that an "intermediate version" — deepseek-v4.1-flash-expires-on-0910 — was open for limited testing. The trailing 0910 is the whole story: this build is scheduled to go offline around September 10.

What DeepSeek actually claimed

The official wording is short. V4.1 Flash is an intermediate checkpoint that uses a new model architecture with native multimodal support, and DeepSeek claims it is stronger, faster and cheaper than V4 Flash. That is a vendor claim in a community-group post accompanied by a feedback form, not a benchmark table — there are no published scores to check it against yet.

Note what is not claimed: nothing about V4 Pro, nothing about the consumer app, and no new price. If you were waiting for the official V4-Pro release promised when V4-Flash went GA on July 31, this is not it.

Why "native multimodal" is the real change

Until now, mixing images and text on the Flash line meant reaching for a separate vision model — deepseek-v4-flash-vision-exp. That approach runs the image through its own encoder and splices the result into the text sequence. It works, but the conversion step costs latency and loses information at the seam.

Native multimodality means images and text are trained and served inside the same architecture. On a product line whose entire reason to exist is low latency, that is a directional decision rather than a feature bullet: it suggests the cheap tier, not the premium tier, is where DeepSeek wants vision to live. If it holds, "add an image to your existing Flash call" becomes the default rather than an experimental detour.

How to test it before it disappears

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_DEEPSEEK_API_KEY",
    base_url="https://api.deepseek.com",
)

resp = client.chat.completions.create(
    model="deepseek-v4.1-flash-expires-on-0910",  # expires 2026-09-10
    messages=[{"role": "user", "content": "Summarise this in one line."}],
)
print(resp.choices[0].message.content)

Three things worth measuring while the window is open, because nobody has published them yet:

• Time to first token vs deepseek-v4-flash on identical prompts, single-threaded, so the 20-request cap cannot distort the result.
• Image-token accounting. Vision-Exp bills images at a capped token budget; check what the new architecture reports in usage for the same image.
• Cache-hit behaviour on repeated prefixes — cache pricing is where the Flash tier gets genuinely cheap, and a new architecture can change hit rates.

What it signals about pricing

Billing during the beta matched V4 Flash exactly, so nothing changed on invoices at the time. But "lower cost" was one of the three claims, and DeepSeek has form here: the V4-Pro rate cut became permanent, and the long-announced peak-hour surcharge went live with the GA release on September 9, 2026 — off-peak rates are now half of peak. Our DeepSeek pricing page tracks the official rate card and is checked weekly.

The read at the time: an architecture preview, priced as its predecessor to remove any reason not to test it. The rate card that followed two days later confirmed the new tiering.

Frequently asked questions