DeepSeek V4.1 Flash: 552B Backbone, MIT Weights and Benchmarks

A new architecture ships as API and weights

DeepSeek released V4.1 Flash through its hosted API as deepseek-flash and published the model repository and weights under the MIT license. It accepts text and images, produces text and tool calls, and supports contexts up to one million tokens. This replaces the April V4 preview rather than merely adding another post-training date to the same design. [1][3]

The stated 552 billion parameters are the Transformer backbone. The model card separately lists a 196-billion-parameter Engram conditional-memory component. Mixture-of-experts routing means only a fraction runs for each token: 8 billion parameters during input processing and 16 billion while generating output. [3]

MIT-licensed weights make the model inspectable and self-hostable, but not small. The official serving examples assume GPU-backed inference frameworks, and DeepSeek's launch page separately invites operators planning deployments of at least 2,000 GPUs plus storage to contact the company. Open access to the weights and easy local deployment are different claims. [1][3]

Why the cache shrank

A model's key-value cache preserves attention state for earlier tokens so it does not rebuild that state every time it generates the next token. Long agent sessions make this cache a large memory and storage cost. V4.1 Flash splits 40 Transformer layers into a 20-layer causal encoder and a 20-layer decoder, then constructs the decoder's global cache from the encoder's final representations instead of keeping a separate full-history cache for every decoder layer. [3]

Compressed Sparse Attention 2 lets layers share cache contents and, in some modes, reuse earlier selections of relevant tokens. The main cache uses four-bit values, while a bounded replay technique reconstructs short-window state instead of persisting all of it. DeepSeek reports 890 bytes of global cache per token, about one quarter of V4 Flash's high-bandwidth-memory footprint, and roughly one eighth of its persistent SSD requirement. [1][3]

Those ratios describe DeepSeek's serving architecture, not a promise that every user's bill falls by three quarters. In Artificial Analysis's comparison, the blended prices were close, about $0.184 per million tokens for V4.1 Flash and $0.177 for V4 Pro, while Flash cost $0.27 per evaluation task versus $0.12 for Pro amid both higher output-token volume and a higher output-token price. Memory efficiency and cost per accepted result are different measures. [1][4]

The benchmark claim is narrower than the headline

DeepSeek's announcement says testing put V4.1 Flash ahead of V4 Pro on performance, cost, speed, and total runtime. Its detailed model card supports the agent-work part of that claim: at maximum reasoning effort, V4.1 Flash led V4 Pro on Terminal-Bench 2.1, DeepSWE v1.1, CyberGym, HLE with tools, and AutomationBench in DeepSeek's comparisons. [1][3]

Across the model card's instruct and base-model tables, Flash does not make a clean sweep. V4 Pro scored higher on GPQA Diamond, the full Humanity's Last Exam result, and several base-model knowledge and long-context tests. The instruct results used reasoning effort 100 and temperature 1.0; the agentic evaluations additionally used benchmark-specific scaffolds. [3]

Artificial Analysis independently measured the hosted max-effort model at 40 on its Intelligence Index, versus 31 for V4 Pro, and about 220 generated tokens per second versus 65. It also recorded about 89,000 output tokens per task for Flash and 49,000 for Pro, so higher generation speed did not make the model terse. [4]

Even under one checkpoint, the surrounding agent matters. DeepSeek's own test resolved 65.5 to 74.2 percent of DeepSWE v1.1 tasks depending on the scaffold, an 8.7-point range without changing the weights. The vendor table and the independent index both support a capable Flash release; neither makes every benchmark or workload equivalent. [3][4]

A model ID changes under its callers

DeepSeek retired V4 Flash and V4 Flash Vision Exp at launch, leaving their old names as temporary aliases to V4.1 Flash. It also announced that deepseek-v4-pro requests would begin routing to V4.1 Flash at 04:00 UTC on September 14 and use Flash rates until V4.1 Pro arrives. That future routing had not happened as of the September 12 check. [1]

The switch turns an unchanged API string into a moving dependency. A developer who relies on V4 Pro's particular reasoning or output behavior must test V4.1 Flash before the routing date, even if the lower rate and smaller cache make the replacement attractive on paper. [1]