DeepSeek V4.1 Flash Beta: Model ID and Expiry

Introduction

Beta closed. Call GA with deepseek-flash — How to call DeepSeek V4.1 Flash officially. This page keeps the 8–10 Sep window; do not ship the temp ID.

Search DeepSeek V4.1 Flash, and many people treat it as another GA release. It is not.

From 8 Sep 2026, a time-boxed model ID became callable in the API environment: deepseek-v4.1-flash-expires-on-0910. Public notes frame it as an intermediate beta: new model structure, native multimodal, faster, same-tier cost. This site is an independent community — as of writing, the official API docs model table still lists deepseek-v4-flash / deepseek-v4-pro / deepseek-v4-flash-vision-exp and has not added V4.1 to a formal Change Log. Below are checkable call facts and limits so you can evaluate inside the window — not wire a temp endpoint into production.

I. What the DeepSeek V4.1 Flash Beta Is — and Is Not

Claim	Fact
GA / permanent model name	No. The ID embeds expires-on-0910 — literally about 2026-09-10
Change base_url	No. Keep your existing DeepSeek API endpoint
New price tier	No. Public notes: billing matches deepseek-v4-flash
High-concurrency production tier	No. About 20 concurrent streams per account (far below shipping Flash)
Stable model listed in docs	No. As of writing, the public model table does not list this ID

In one line: a short evaluation endpoint, not a long-lived production model string. Flagship vs daily routing still follows What is DeepSeek V4 Pro; shipping Flash/Pro rates: DeepSeek API price change.

Vendor claims (no separate whitepaper at writing time): new structure, native multimodal, stronger, faster, lower cost. Community reports often cite higher output throughput than day-to-day Flash — treat speed as something you measure yourself, not an SLA.

II. How to Call: Change Only the Model Name

base_url stays the same — e.g. OpenAI-compatible https://api.deepseek.com. Set model to:

deepseek-v4.1-flash-expires-on-0910

Minimal example:

from openai import OpenAI

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

response = client.chat.completions.create(
    model="deepseek-v4.1-flash-expires-on-0910",
    messages=[{"role": "user", "content": "Explain in three sentences what you are good at."}],
)

print(response.choices[0].message.content)

Keys still come from the official platform. API overview: DeepSeek API; from-zero steps: DeepSeek web & API beginner guide.

Must do:

1. Keep this ID in eval / sandbox configs — not production default routes.
2. Calendar reminder: around 2026-09-10, switch back to deepseek-v4-flash or deepseek-v4-pro.
3. Watch concurrency (~20); saturation hits rate limits — backoff: DeepSeek API 429.
4. Reconcile bills against the Flash tier; peak/off-peak and cache still follow shipping price logic: context caching.

III. vs V4 Flash, Vision-Exp, and V4 Pro

| Model / entry | What you get | Use it now for | | --- | --- | | deepseek-v4-flash | Shipping Flash text tier (in docs) | Daily throughput, production default | | deepseek-v4-flash-vision-exp | Experimental vision add-on on Flash | Image input already; accept -exp instability | | deepseek-v4.1-flash-expires-on-0910 | V4.1 intermediate beta, native-multimodal wording + short window | ~48h A/B eval only — never as default | | deepseek-v4-pro | Shipping flagship | Long agents, hard reasoning, full vision path |

vs Vision-Exp (public wording): Vision-Exp looks like vision bolted onto Flash; V4.1 Flash beta stresses structure-level native multimodal. Capability edges depend on your own image+text cases — no separate V4.1 scorecard in official docs at writing time.

Some will ask whether it can replace Pro — that is a vendor survey direction, not a conclusion. Until docs update, production routing should stay: default Flash/Pro; beta ID for controlled experiments only.

IV. Eval Checklist (Short Window)

Finish these before expiry — do not “wire first, measure later”:

1. Latency & throughput: same prompts vs deepseek-v4-flash; log tokens/s (self-measured, not marketing numbers).
2. Multimodal: at least one screenshot OCR and one chart read; same images on Vision-Exp.
3. Agent / tool loops: with thinking on, respect replay rules — DeepSeek thinking mode.
4. Rate limits: push near 20 concurrent; confirm your retry policy.
5. Failback drill: one-click switch back to shipping Flash so 10 Sep does not blank production.

V. FAQ

Is this an official GA launch?
Per public docs: the model table and Change Log did not list it at writing time. Treat it as a callable short beta endpoint — do not write “official GA” copy.

Does it still work after 10 Sep?
By the ID literal, that string expires. Trust the platform response then; default assumption: switch back to shipping model names.

Is it cheaper?
Public notes: same price as deepseek-v4-flash. “Lower cost” is mostly same-price capability/speed feel — not a new rate card.

Can I pick V4.1 on the web app?
This article covers the API model-name path only. Whether the consumer client exposes the tier depends on the product UI.

Summary

Shortest DeepSeek V4.1 Flash beta path: set model to deepseek-v4.1-flash-expires-on-0910 → sandbox eval at Flash price and ~20 concurrency → before 2026-09-10, cut back to Flash/Pro listed in docs. New structure and native multimodal are worth testing; a temp ID is not worth production defaults.

Further reading

• DeepSeek V4.1 Flash
• DeepSeek API
• What is DeepSeek V4 Pro
• DeepSeek API price change
• DeepSeek API 429 rate limit
• DeepSeek thinking mode
• DeepSeek web & API beginner guide