New Released Sep 10, 2026 · DeepSeek V4.1 Flash

Flagship intelligence, at a small model’s cost

DeepSeek V4.1 Flash is a 552B MoE model with an asymmetric Causal-Encoder-Decoder architecture: only 8B parameters are active on the input side and 16B on the output side. A 1M-token context, native image understanding, and compatibility with the OpenAI and Anthropic APIs — you only need to rename the model to deepseek-flash.

deepseek-flash
quickstart.sh
curl https://api.deepseek.com/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $DEEPSEEK_API_KEY" \
  -d '{
        "model": "deepseek-flash",
        "messages": [
          {"role": "user", "content": "Hello!"}
        ],
        "stream": false
      }'
1M
Context
384K
Max output
$0.30 / $0.15
/1M input tokens (peak / off-peak)
Total parameters
552B
MoE, CED architecture
Active on input
8B
prefill
Active on output
16B
decode
KV cache
1/4
HBM demand vs. the previous generation
Concurrency
2,500
concurrency limit
Off-peak window
−50%
vs. peak pricing

Overview

A new architecture generation, starting with the smallest model

V4.1 Flash is the smallest model in DeepSeek’s brand-new architecture line. The design philosophy is clear: raise the capability ceiling, speed up reasoning, increase throughput, and open a path to scale up to larger parameter models. As a result, a low-cost model beats the previous generation of flagship models on many agentic tests.

Asymmetric design

Input and output use different activation levels (8B / 16B), delivering higher capability at a far lower cost than other models of a similar size.

Smaller cache, cheaper agents

The KV cache is compressed aggressively: HBM demand drops to 1/4 and SSD to 1/8 versus the previous generation. For agent workloads — where cache hits make up a large share of the bill — this is the decisive factor.

Native multimodal

Image understanding is built into the model (native vision) from the very first Flash release — something deepseek-v4-pro does not have.

Architecture

Causal-Encoder-Decoder

Instead of a single computation path for both reading and writing, CED splits the work into two phases with different costs: the read phase (prefill) activates only 8B parameters, while the generation phase (decode) activates 16B. Combined with new pre-training and large-scale RL post-training, the model reaches reasoning levels that surpass its own house’s flagships.

  • Balanced read/write cost — most long prompts only cost the 8B prefill.
  • Built to scale up — the design is meant to grow into larger parameter models on the same architecture.
  • Optimized for agents — a smaller KV cache fits more context and more parallel sessions on the same GPU cluster.

Full technical details are in the DeepSeek V4.1 technical report .

Prompt / context (up to 1M tokens)
Text · Images · Tool schema
Encoder · Prefill
8B
active parameters for input
Decoder · Decode
16B
active parameters for output
Output
Text · JSON · Tool calls · Reasoning trace (up to 384K tokens)
552B
total MoE parameters
−75%
HBM demand for the KV cache
437×
smaller KV cache than the first generation

Features

Everything production workloads need

Model capabilities come with a full API surface: tool calling, structured output, reasoning modes, vision, and the two most widely used interface standards today.

Dual reasoning modes

Supports both thinking and non-thinking. Thinking is on by default; turn it off for fast, cheap responses on simple tasks.

Tool Calls & JSON Output

Call tools and constrain the output to JSON — a good fit for pipelines that need structured data rather than free-form text.

Native vision

Read images directly in the same API call. No OCR pipeline or separate vision model required.

1M context · 384K output

Fit an entire codebase or long document into a single call, and get answers at the scale of hundreds of thousands of tokens.

OpenAI & Anthropic compatible

Two separate base URLs, so you can reuse existing SDKs, agent frameworks, and tooling. Moving to Flash means changing one line: the model name.

Responses API, FIM & prefix

A native Responses API for modern agents, plus FIM completion and chat prefix completion (beta, only in non-thinking mode).

Performance

Flash outperforms Pro

According to DeepSeek’s announcement, V4.1 Flash has surpassed several flagship models — including DeepSeek V4 Pro — on agentic benchmarks, while also doing better on cost, speed, and total processing time. That is why DeepSeek is retiring V4 Pro from the API and steering users toward Flash.

A note on the numbers: detailed benchmark figures are in DeepSeek’s technical report and official comparison tables. This page does not invent scores — compare directly with the original sources.

Comparison of deepseek-flash and deepseek-v4-pro
Criteria deepseek-flash deepseek-v4-pro
Model version DeepSeek-V4.1-Flash DeepSeek-V4-Pro-0813
Context / output 1M / 384K 1M / 384K
Vision Yes No
Concurrency limit 2,500 500
Input price (cache miss) / 1M $0.30 peak $1.32
Input price (cache hit) / 1M $0.006 peak $0.044
Output price / 1M $1.20 peak $3.96
Status Recommended Being retired

Peak pricing is used for direct comparison; off-peak pricing is half. Figures from DeepSeek’s Models & Pricing page. Prices may change without notice.

Pricing

Per-token pricing, with off-peak rates

Priced per million tokens, in USD. DeepSeek uses time-of-day pricing: outside peak hours you pay only half the rate.

Price per 1 million tokens
Token type (per 1M) deepseek-flash deepseek-v4-pro
Input — cache hit repeated prompt already in cache $0.006 off-peak $0.003 $0.044 off-peak $0.022
Input — cache miss new content never sent before $0.30 off-peak $0.15 $1.32 off-peak $0.66
Output includes thinking tokens $1.20 off-peak $0.60 $3.96 off-peak $1.98

Peak hours (Vietnam time, UTC+7): 08:00–11:00 and 13:00–17:00 Monday – Friday (excluding Chinese public holidays) — that is 01:00–04:00 and 06:00–10:00 UTC. Every other window — including full weekends and holidays — is off-peak.

Estimated monthly cost

Adjust it to match your actual traffic.

deepseek-flash —
deepseek-v4-pro —

Get started

Integrate in minutes

Get an API key from the DeepSeek Platform, pick the base URL for the standard you already use, and set model: "deepseek-flash". No new SDK to learn.

Base URL · OpenAI-compatible
https://api.deepseek.com
Base URL · Anthropic-compatible
https://api.deepseek.com/anthropic
curl https://api.deepseek.com/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $DEEPSEEK_API_KEY" \
  -d '{
        "model": "deepseek-flash",
        "messages": [
          {"role": "system", "content": "You are a technical assistant. Answer concisely."},
          {"role": "user", "content": "Explain the Causal-Encoder-Decoder architecture in 3 bullet points."}
        ],
        "stream": false
      }'

The examples illustrate syntax only; for details on parameters (thinking mode, tool calls, prefix completion) see the DeepSeek API Docs.

Roadmap

From V4 preview to V4.1 Flash

  1. Apr 24, 2026
    V4 preview

    Two variants, V4-Pro and V4-Flash, launched via API.

  2. Aug 13, 2026
    V4-Pro general availability

    The official release of the V4-Pro flagship model.

  3. Aug 21, 2026
    V4-Flash-Vision-Exp

    Multimodal API opened for the Flash line.

  4. Sep 10, 2026
    V4.1 Flash launch

    CED architecture, model ID deepseek-flash, new pricing in effect, weights released.

  5. Sep 14, 2026
    V4 Pro migration

    From 12:00 Beijing time, requests to deepseek-v4-pro are served by V4.1 Flash.

Official partners

Tencent (WorkBuddy, CodeBuddy) and OpenCode have fully integrated V4.1 Flash.

Open weights

deepseek-ai/DeepSeek-V4.1-Flash on Hugging Face, together with the technical report.

Large-scale deployment

DeepSeek helps teams with suitable infrastructure (around 2k GPUs plus a storage cluster) run the model themselves.

FAQ

Frequently asked questions

Can’t find the answer you need? Look it up in the API docs or the official FAQ.

How is deepseek-flash different from deepseek-v4-flash?

deepseek-flash is the new name for DeepSeek-V4.1-Flash. The older models deepseek-v4-flash and deepseek-v4-flash-vision-exp have been retired, but they still work for compatibility reasons — such requests are served by V4.1-Flash and billed at Flash rates.

Do I have to rewrite code to move to Flash?

No. Just change the model name in your payload and point the base URL at https://api.deepseek.com (OpenAI-compatible) or /anthropic. SDKs, message structure, and tool calls stay the same.

How do I turn thinking mode on or off?

V4.1 Flash supports both modes, and thinking is the default. Non-thinking mode is for tasks that need low latency and costs less; it is also the only mode that supports FIM completion. See the Thinking Mode section of the docs for how to switch.

When are off-peak hours?

Measured in Vietnam time (UTC+7), peak hours are 08:00–11:00 and 13:00–17:00, Monday through Friday, excluding Chinese public holidays (01:00–04:00 and 06:00–10:00 UTC). Outside those windows — including full weekends and holidays — prices are halved. For teams in Vietnam, the cheapest hours of the day are the evening and overnight (after 17:00 until 08:00 the next day), plus the 11:00–13:00 lunch break.

What happens to deepseek-v4-pro?

DeepSeek is retiring V4 Pro. From 12:00 on Sep 14, 2026 (Beijing time) until V4.1 Pro launches, every request to deepseek-v4-pro is served by V4.1 Flash and billed at Flash rates.

Can I self-host the model?

The weights are published publicly on Hugging Face. For large-scale needs, DeepSeek invites you to get in touch if you have the corresponding infrastructure (around 2k GPUs plus a storage cluster).

Change one line of model name, cut your bill sharply

Flagship capability at small-model pricing. Start with a 1M-token context, native vision, and an API compatible with the OpenAI / Anthropic standards.