DeepSeek V4.1 Flash: Architecture, Benchmarks, Pricing, and Migration Guide

DeepSeek-V4.1-Flash is a general-availability multimodal model released on September 10, 2026. It is the first production model in DeepSeek’s Causal Encoder–Decoder (CED) family: a 552-billion-parameter Mixture-of-Experts model designed to make long prompts, cached context, vision input, and agent workflows more economical. 416

The practical takeaway is straightforward: V4.1 Flash is positioned as DeepSeek’s default high-throughput API model for large-context and multimodal work. Its reported advantages over V4 Pro are strongest in cost, memory efficiency, and several agent-oriented tests; its disclosed scores do not show it winning every benchmark. 2411

What the 8B/16B architecture changes

A conventional shorthand such as “552B parameters” does not describe how much of a sparse MoE model is used for every token. DeepSeek says V4.1 Flash activates roughly 8B parameters during input prefill—when it reads a prompt—and 16B during decoding, when it generates tokens. 4138

That asymmetric design matters for applications that repeatedly feed the model large documents, repositories, conversation histories, or tool traces. Prompt ingestion and answer generation have different computational demands, so the CED architecture is intended to reduce the cost of reading large contexts while retaining more compute for output generation. 38

DeepSeek also reports a global KV-cache footprint of about 890 bytes per token, approximately one quarter of the prior V4 Flash figure. It attributes this to FP4 KV caching and cross-layer KV-cache reuse. For persistent agents, where cached context can become a major infrastructure cost, this may be as consequential as the headline parameter count. The figure is a provider/model-card claim and should be validated against a deployment’s own serving stack. 3338

Capability versus V4 Pro: the published results

DeepSeek’s published table shows gains in programming, mathematics, terminal-agent, and cyber benchmarks relative to V4 Pro, but a lower GPQA Diamond score. These are reported/model-card figures rather than an independent universal ranking. 611

Benchmark	V4 Pro	V4.1 Flash	Difference
GPQA Diamond	92.4	90.9	-1.5
Codeforces rating	3,348	3,471	+123
MathArena Apex	65.3	65.6	+0.3
Terminal-Bench 2.1	87.9	90.6	+2.7
CyberGym	83.3	88.1	+4.8

911

The fairest reading is therefore nuanced. V4.1 Flash looks better suited to DeepSeek’s stated product goal—fast, cost-conscious, long-running agent and coding workflows—while V4 Pro remains ahead in the disclosed GPQA Diamond result. A claim that Flash is categorically superior at every task is not supported by the published comparison. 11

DeepSeek has said that testing by multiple parties put V4.1 Flash ahead of V4 Pro in performance, cost, speed, and total runtime. However, the provided public materials do not establish a single independently audited multiplier for end-to-end agent task time. Treat that broader comparison as a vendor claim, then benchmark it against the tools, context lengths, and task mix that matter in production. 24

Context, vision, and API capabilities

V4.1 Flash supports up to 1 million tokens of context and accepts text and images in the same request. DeepSeek documents image description, screenshot text reading, and chart analysis as example uses. 623

For the API, the current model identifier is:

deepseek-flash

Developers can provide a publicly accessible http(s) image URL alongside text. DeepSeek’s vision documentation lists JPEG, PNG, GIF, and WebP as supported formats. 2123

The supplied documentation establishes the 1M-token context limit, but it does not provide a sufficiently clear current maximum output-token allowance to state one here. Likewise, developers should verify the live API schema for the precise availability and syntax of JSON output, tool calling, and reasoning controls before relying on a specific integration behavior.

Pricing and the cache-hit incentive

Reported V4.1 Flash API prices are $0.15 per million input tokens and $0.60 per million output tokens off-peak, rising to $0.30 input and $1.20 output at peak. Reported cached-input prices are $0.003 off-peak and $0.006 peak per million tokens. 1

That unusually low cache-hit price fits the architecture’s focus on repeated context. Applications that reuse a large system prompt, codebase context, or agent history could see more benefit than one-off short chats—assuming their request patterns produce cache hits.

Migration: what happened to the old models?

DeepSeek retired V4 Flash and V4 Flash Vision Experimental as separate models. For compatibility, the legacy identifiers below remain accepted temporarily and are served by V4.1 Flash:

• deepseek-v4-flash
• deepseek-v4-flash-vision-exp

2123

DeepSeek’s original September 10 announcement said deepseek-v4-pro traffic would route to V4.1 Flash at Flash rates from 04:00 UTC on September 14, pending the arrival of V4.1 Pro. 24

Later API documentation revised that operational picture: it says DeepSeek would continue providing V4 Pro API service after September 14 and keep its billing unchanged, with further notice if that changes. Teams using V4 Pro should rely on the current API documentation and test the model actually returned by their endpoint rather than assume the earlier retirement plan remains in force. 1726

Why this is a product-line reset

V4.1 Flash is more than a renamed Flash endpoint because several changes arrived together:

• a new Causal Encoder–Decoder MoE architecture with different input and output activation levels; 41
• native multimodal support in the primary deepseek-flash endpoint; 2423
• a 1M-token context limit and a substantially smaller reported KV-cache footprint; 638
• retirement of the previous Flash and Flash Vision Experimental models, with compatibility aliases maintained during migration. 2123

For developers, the sensible path is to move new integrations to deepseek-flash, test image handling and long-context behavior on representative workloads, measure cache-hit economics, and keep an eye on the live V4 Pro routing and billing documentation. The architecture is promising for agent-heavy systems, but the relevant proof is performance and cost in the specific workflow being deployed.