DeepSeek-V4.1-Flash

DeepSeek-V4.1-Flash is DeepSeek's most efficient frontier model, released on September 10, 2026. It is a 552B-parameter Mixture-of-Experts model built on a new Causal Encoder–Decoder architecture — 8B active parameters for input and 16B for output — with native vision, a 1M-token context window, and a KV cache that needs one quarter of the HBM and one eighth of the SSD storage of the previous generation.

It replaces DeepSeek V4-Flash, which has been retired: the legacy deepseek-v4-flash API name is still accepted and now routes to V4.1-Flash. DeepSeek also plans to route all deepseek-v4-pro requests to V4.1-Flash starting September 14, 2026 (12:00 Beijing time) until a future V4.1-Pro release.

DeepSeek rebuilt the model instead of iterating on V4-Flash. The Causal Encoder–Decoder design splits input reading (8B active parameters) from output generation (16B), and DeepSeek says new pretraining methods plus larger-scale RL post-training push benchmark results ahead of its own V4-Pro flagship. The storage profile changed with it: shrinking the KV cache to a quarter of the HBM and an eighth of the SSD footprint mainly lowers the cache-hit cost of long agent sessions, where cached context is often the largest line item.

Prices dropped with the launch, and off-peak rates are half of peak rates (peak hours: 01:00–04:00 and 06:00–10:00 UTC, Monday–Friday), so flexible batch work can be scheduled more cheaply.

• Long-context efficiency — a 1M-token context window with a KV cache designed to cut memory and storage cost.
• Native vision — image input is supported natively alongside text.
• Agent-ready API — tool calls, JSON output, the Responses API, and an Anthropic-format endpoint.
• Thinking mode — supports both thinking (default) and non-thinking modes.
• Continuity — legacy deepseek-v4-flash requests keep working at Flash prices during the transition.

For deeper analysis — architecture details, pricing mechanics, and benchmark interpretation — see our DeepSeek V4.1 Flash blog. Verify current pricing, limits, and availability in DeepSeek's official documentation before relying on them, because the provider can change the rate card and model status.