DeepSeek V4.1 Flash Launches with Major VRAM Reduction: what are the key facts?

DeepSeek has released the V4.1 Flash model, featuring new CED and CSA2 architectures.

DeepSeek V4.1 Flash Launches with Major VRAM Reduction: what are the key facts?

Runtime VRAM usage has decreased by 75%, while persistent cache is reduced to one-eighth of the previous generation.

DeepSeek V4.1 Flash Launches with Major VRAM Reduction: what are the key facts?

The model achieved a 74.2% pass rate on the DeepSWE benchmark, outperforming several competitors.

DeepSeek V4.1 Flash Launches with Major VRAM Reduction: what are the key facts?

With 552B total parameters, the model operates using only 8B active parameters during the prefill stage.

Physix Frontier · News Briefing Card (enbrief · Sep 11, 2026)

DeepSeek V4.1 Flash Launches with Major VRAM Reduction

KEY FACTS

DeepSeek has released the V4.1 Flash model, featuring new CED and CSA2 architectures.
Runtime VRAM usage has decreased by 75%, while persistent cache is reduced to one-eighth of the previous generation.
The model achieved a 74.2% pass rate on the DeepSWE benchmark, outperforming several competitors.
With 552B total parameters, the model operates using only 8B active parameters during the prefill stage.

KEY DATA

75%VRAM Usage Reduction

1/8Persistent Cache Ratio

74.2%DeepSWE Pass Rate

552BTotal Parameters

PHYSIX OBSERVATION

By prioritizing architectural innovation over brute-force scaling, DeepSeek has broken the high-cost curse of long-context agents. The cliff-like drop in VRAM and storage costs signals that million-token applications are moving from the lab to large-scale commercial deployment. This 'small bet big' sparse pathway may force the industry to reevaluate the cost-efficiency logic of dense LLMs, accelerating the adoption of low-cost intelligent agents.

Source: enbrief original report ↗