DeepSeek V4.1 Flash Takes Over for Pro: Asymmetric Design Joins the Frontier-Agent Race

DeepSeek broke an unwritten rule with this release: the model called Flash is temporarily taking over work that used to go to the flagship Pro. Starting 04:00 UTC on September 14, 2026, Pro requests route to V4.1 Flash at Flash rates until the next Pro model lands. That is not a side option for cheap traffic; it is the main line wearing a Flash badge for a while.

The old hierarchy was simple. V4 Pro was the heavyweight at 1.6 trillion parameters with 49 billion active per token, the model you picked for maximum strength. V4.1 Flash arrives with a 552-billion-parameter backbone, which is smaller yet still enormous. The headline is therefore not the parameter count but how much of the model wakes up at each stage of the work.

The core idea is an asymmetric Causal Encoder-Decoder split. Roughly speaking, reading and writing are handled by different crews: about 8 billion active parameters organize the prompt, the codebase, and tool outputs while reading, then about 16 billion take over when the model writes answers and picks the next tool call. The mental image is one team sorting a huge case file and handing the distilled pieces to a stronger team that drafts the plan.

Memory is where the design pays rent. Long agent runs accumulate a growing record of processed context, and carrying that record gets expensive as windows stretch toward a million tokens. DeepSeek quotes around 890 bytes of global cache per token, about one quarter of the previous generation in active memory and one eighth in persistent storage. Smaller cache means less high-bandwidth memory per live run, less SSD per stored session, and less repeated work when the agent revisits earlier material.

Then come the tables. On Terminal-Bench 2.1 the video reports 90.6 for Flash against 89.1 for Claude Opus 5 and 88.8 for GPT-5.6 Sol; on DeepSWE it reports 74.2 against 74.0 and 73.0; on AutomationBench the gap widens to 54.8 against 50.3 and 45.8. Early independent chatter adds texture: one intelligence index scores Flash at 40 points above its own Pro at 36, with domestic rivals higher still, and a design-arena test puts it within a point and a half of a much pricier closed model at a fraction of the cost per finished design.

The model does not work alone, which is why the Harness update matters. DeepSeek frames the Harness as everything around the brain: tools, files, shell access, planning, subagents, sessions, storage, scheduling. The new version keeps expensive context alive when system instructions change mid-task, exposes cache-hit and token statistics more clearly, and refreshes optional subagent runtimes. Version 0.1.5 also leans into file uploads and sidebar previews, pushing the Harness from a runner toward a daily workbench.

Open weights plus MIT licensing round out the picture, with a catch the video states plainly. Developers gain inspection, supported inference stacks, and freedom from a single closed API, but the full backbone still has to live somewhere far beyond a standard laptop. So the decision splits three ways: API builders get a cheaper reader for long agents, infrastructure teams get control at the price of serious hardware, and anyone chasing a single champion for every task gets evidence of membership in the frontier conversation rather than proof of a crown. The thesis of this release is asymmetry: be cheap where reading allows it and strong where deciding demands it.