DeepSeek-V4.1-Flash is a mixture-of-experts model released by DeepSeek on September 9, 2026. It succeeds the V4 family of April 2026 and replaces V4-Pro as the company's served flagship pending a V4.1-Pro release.
Everything recorded here comes from DeepSeek's own announcement. No technical report has been examined, no independent benchmark has been located, and no third-party evaluation is held, so the specification and efficiency claims below are the developer's and are unverified.
Architecture and specifications
The model carries 552 billion total parameters on what DeepSeek describes as a new causal encoder–decoder architecture, activating 8 billion parameters for input and 16 billion for output. The asymmetry between input and output activation is a departure from the uniform per-token activation of the V3 and V4 lines, which activated 37 billion and 49 billion parameters respectively regardless of position. It adds native visual understanding, which the V4 family did not carry.
DeepSeek states that the model's key-value cache requires one quarter the high-bandwidth memory and one eighth the solid-state storage of the previous generation (Source: deepseek.com).
Availability and pricing
New API pricing took effect at 04:00 UTC on September 10, 2026, with off-peak rates set at 50% of peak — continuing the time-of-day pricing DeepSeek introduced across the V4 line in August 2026. From 04:00 UTC on September 14, 2026 all deepseek-v4-pro API requests route to V4.1-Flash at V4.1-Flash rates until V4.1-Pro launches (Source: deepseek.com).
Routing a flagship model's traffic to a smaller successor is a substitution rather than a deprecation notice with a migration window, and the announcement gives no date for V4.1-Pro.
Relation to the compute constraint
The memory claim is the part of the release with policy relevance. High-bandwidth memory supply is assessed as the binding constraint on Chinese AI compute production through roughly 2028, with domestic supply expected to narrow but not close the gap (Will Huawei catch up to Nvidia by 2030? (Epoch AI, September 2026)). An architecture that reduces cache memory per served request by a factor of four eases that constraint on the inference side, where Chinese domestic silicon is already assessed as more competitive than in training (Huawei — Ascend AI Accelerators).
The V4 line disclosed Huawei Ascend training compatibility as a deliberate design constraint. Whether V4.1-Flash carries the same compatibility is not stated in the announcement.
Open questions
- Whether the weights have been released, and under what licence, is not established from the announcement.
- No benchmark results, safety evaluations or system card have been located.
- Whether the causal encoder–decoder architecture and the input/output activation asymmetry are described in a technical report, and how they produce the stated cache reduction, is unresolved.
- Whether the model trains or serves on domestic Chinese accelerators is not stated.
Relationships
- supersedes: DeepSeek V4 Pro / V4 Flash — replaces V4-Pro on the API from September 14, 2026
- related: DeepSeek — developer
- related: Inference Economics and Token Pricing — the memory and pricing claims
- related: Will Huawei catch up to Nvidia by 2030? (Epoch AI, September 2026) — the memory constraint the efficiency claim bears on
- depends-on: DeepSeek-V4.1-Flash Technical Report (DeepSeek, September 2026) — DeepSeek's technical report for the model