Skip to content

Model detail

deepseek-v4.1-flash

Provided by DeepSeek
Pay-as-you-go

DeepSeek V4.1 Flash is a sparse mixture-of-experts model from DeepSeek, and the first built on the company's Causal Encoder-Decoder (CED) architecture. It activates 8B parameters on input and 16B on output from a 552B-parameter backbone, an asymmetric split that keeps per-token compute low relative to the model's total size. Image understanding is native to the architecture, with visual and text embeddings trained jointly from the start of pre-training rather than added afterward as in the earlier experimental [V4 Flash Vision Exp] It is suited for coding, terminal, and computer-use agents, along with long-horizon tasks that must run to completion across many steps and long-context analysis. Compressed KV caching cuts cache memory to roughly a quarter of the previous Flash generation, significantly reducing costs on agentic workloads. DeepSeek positions it as the cost-efficient tier of the V4.1 family and reports that it exceeds [V4 Pro] on performance, speed, and task completion time.

Model specs

Model ID
deepseek-v4.1-flash
Context length
1M
Max output
384K
I/O modalities
Released
2026-09

02

API endpoints

One MixRoute gateway, OpenAI-compatible

  • OpenAI-compatible /v1/chat/completions POST

03

Pricing

Flat rate, unit: /1M Tokens

Input

$0.3000 /1M Tokens

Completion

$1.2000 /1M Tokens

Cache read

$0.0060 /1M Tokens

05

Selection summary

Quickly judge whether it fits your workload

What is this model good for?

Sparse MoE model for agentic and long-horizon work

DeepSeek V4.1 Flash is a sparse mixture-of-experts model on DeepSeek's Causal Encoder-Decoder architecture, activating 8B parameters on input and 16B on output from a 552B backbone with native image understanding. It suits coding, terminal and computer-use agents, long-horizon tasks and long-context analysis.

What should you check before using it?

Check tier position, context limits and cache savings

Confirm your context and output limits, and note two design points: compressed KV caching cuts cache memory to roughly a quarter of the previous Flash generation, and DeepSeek reports this tier exceeds V4 Pro on speed and task completion. Validate both on your own agent traces. Knowledge cutoff: 2025-05.

Why use it through MixRoute?

Cache reads at 0.006 USD per 1M through MixRoute

MixRoute bills cached prefixes at 0.006 USD per 1M tokens against 0.30 USD input and 1.20 USD output, which is where agentic loops save most. Serve it over the OpenAI-compatible endpoint: point your client at https://api.mixroute.ai/v1 and set the model to deepseek-v4.1-flash.

07

Capabilities & compatibility

Unconfirmed capabilities stay "Unknown," no guesses

Available:

Tool calling

DeepSeek-V4.1-Flash supports function and tool calls.

EX Send a tools definition; the response returns tool_calls to execute and feed results back.
Available:

Structured output

DeepSeek-V4.1-Flash can produce JSON-formatted output on demand.

EX response_format: { type: "json_schema" } forces valid JSON output.
Available:

Streaming

DeepSeek-V4.1-Flash supports streamed responses.

EX stream: true streams tokens incrementally, so chat UIs render as they arrive.
Available:

Prompt caching

DeepSeek-V4.1-Flash supports prompt caching under supported routing conditions.

EX Keep the system prompt and tool definitions at the start of the request; repeated prefixes are automatically billed at the cache-read rate.
Available:

Vision input

DeepSeek-V4.1-Flash can process image inputs alongside text.

EX Pass a base64-encoded image in a image_url content part alongside text messages.
Available:

OpenAI SDK

Call DeepSeek-V4.1-Flash through MixRoute's OpenAI-compatible endpoint.

EX Change base_url to https://api.mixroute.ai/v1; everything else follows the OpenAI style.

08

Token cost estimator

Live estimate from this page's pricing, not an actual bill

Share of the same prefix read repeatedly, up to 100%

09

FAQ

DeepSeek-V4.1-Flash is a reasoning model. Reasoning effort can be tuned per request. Lower effort settings cut reasoning tokens and cost for routine work; reserve higher settings for problems that need deeper search. Measure quality at two or three settings before fixing one in production.

One endpoint, a testable decision

Test this model and alternate routes with the same request format

Start from a real workload, then let quality, total cost, and failure conditions decide whether to send production traffic.

Create an API key View integration docs