V4.1 Flash

V4.1 Flash is natively multimodal in design, capable of directly viewing and understanding image content.

This update adds support for the official model ID deepseek-flash; new sessions default to V4.1 Flash, but explicitly configured models take higher priority.

Interfaces that previously called DeepSeek V4 Pro will automatically switch to the new V4.1 Flash model, requiring no code changes from developers.

V4.1 Flash activates only 8 billion parameters during the input stage and 16 billion parameters during the output stage.

DeepSeek open-sourced the inference framework DeepInfer on October 20, providing optimized operators, multi-card parallelism and asynchronous scheduling capabilities for V4.1 Flash, under the Apache 2.0 license.

At 12:00 noon on September 14, V4 Pro will officially be taken offline, and all requests will be fully routed to V4.1 Flash.

The DeepSeek API offers three effort levels—low, high (default) and max—which the V4.1 Flash paper maps to internal effort values of 50, 75 and 100, and reports that going from high to max lengthens agent trajectories by 1.6–1.8 times while yielding only marginal gains.

V4.1 Flash reduces KVCache to one-quarter of the previous generation.

The old model's implementation of the marquee selection feature had a bug: left-clicking first selected the object, and after dragging, the selection was canceled before entering marquee selection.

The suspected V4.1 Flash, gray-tested on October 17, implemented the marquee selection feature for the same prompt "very perfectly", correctly handling the order of click-versus-drag determination.

Whether V4.1 Flash has been officially released and what its specific version number is have not yet been confirmed by the official side.

DeepInfer currently supports only one model, V4.1 Flash, and does not support Hopper and earlier GPUs.

The V4.1 Flash model contains 384 experts, only a few of which are activated per token; these experts are distributed across 64 cards, 6 per card, plus a clone of one popular chef.

V4.1 Flash's expert scheduling strategy: load is monitored every 1,000 steps, experts are reallocated every 10,000 steps, and ultimately the busiest card is only 9% busier than the average level.

September's V4.1 Flash consumed nearly twice as many tokens as open-source models of the same scale when running the same benchmark.

The September version's promotional page said '552B Backbone Parameters', emphasizing backbone parameters; this time the basis was switched from backbone parameters to all weights, and the architecture itself is unchanged.

V4.1 Flash's software engineering score of 74.2 is the highest in the comparison table, but leads the strongest flagship competitor by only 0.2 points.

Migration recommendations: legacy Flash users can upgrade directly; V4 Pro users should conduct migration testing as soon as possible with a focus on long-context numerical memory; users with long-document high-consistency requirements should run a small-scale pilot first.

A community benchmark completed before the September 10 price cut (300 million tokens, 14 task groups, over 2,400 requests) shows that under the same billing basis, V4.1 Flash's total cost is 36% higher than the previous version.

The asymmetric input-output architecture of V4.1 Flash is designed to align with the distribution pattern in Agent tasks where input tokens far outnumber output tokens, thereby reducing overall inference costs.