DeepSeek V4.1.
See more. Build smarter.

Released on 10 September 2026, V4.1-Flash is a multimodal mixture-of-experts model. DeepSeek describes a causal encoder–decoder with 8B active parameters for input and 16B for output. Its launch reports lower KV-cache memory requirements than the previous generation. That is a claim about inference memory, not a promise that your application runs four times faster. DeepSeek launch announcement.

The practical opportunity is a document workflow that needs both the words and the picture: a chart in a report, a screenshot attached to a support case, or a diagram alongside a maintenance note. The model can propose an interpretation; the application still needs to verify important values against the original evidence.

Illustrative calculation: a batch totalling 10 million uncached input tokens and 1 million output tokens costs $4.20 on Flash or $17.16 on Pro at these peak rates. Calculation: input millions × input rate + output millions × output rate. This assumes equal token counts; it excludes tools, retries, infrastructure, tax and human review. Off-peak prices are half the peak rates. Pricing and current peak-hour schedule.

For a real decision, replay the same work on both models. Count rejected answers, additional reasoning, retries and manual corrections. A cheaper answer that takes longer to check can be the more expensive outcome. Our Jev model-router project applies this principle at request level: choose an eligible model, then verify the result.

DeepSeek recommends deepseek-flash. The old V4-Flash and V4-Flash-Vision-Exp aliases now resolve to V4.1-Flash. The current documentation also says V4 Pro remains available after 14 September, superseding the original launch plan to reroute it. Current API migration notice.

Treat an alias change as a model change: replay representative prompts, verify image handling and structured outputs, and compare the returned usage. API-format compatibility does not prove identical tool behaviour. Keep the previous approved route available until the acceptance criteria pass.

Self-hosting is a separate infrastructure project. Active parameters do not tell you how much memory the full deployment needs. Size weights, context cache, concurrency and the serving stack together. Confirm runtime support for this exact release before buying hardware.

Model provenance, provider jurisdiction and data flow are separate decisions. For a hosted service, review the actual processing locations, retention terms and contract. For self-hosting, inspect outbound connections, telemetry, tools and update mechanisms; hosting weights locally alone does not prove that the whole application is isolated.

Pin the model artefact, retain its licence, and evaluate behaviour on the topics your customers actually ask about. Our model supply-chain approach covers that review. If the organisation cannot approve the provider or operate the serving infrastructure, choose another eligible model before comparing scores.