DeepSeek released DeepSeek-V4.1-Flash on September 10, 2026, calling it the smallest model in a new architecture family and the one that will take over V4-Pro traffic on its API. Weights are on Hugging Face under an MIT license. The checkpoint is a 552 billion parameter multimodal mixture-of-experts model with a 1 million token context. It reads images and text and writes text.

The design is a causal encoder-decoder: 20 encoder layers, 20 decoder layers. Prefill activates about 8 billion parameters per token. Decode activates about 16 billion. The lab said that split, plus a compressed KV cache, is aimed at agent jobs that spend more tokens on input than on output. Compared with V4-Flash, it claims one quarter the HBM for the cache and one eighth the SSD. The model card puts the global KV cache at 890 bytes per token.

API cutover and open weights

The API name is deepseek-flash. DeepSeek retired V4-Flash and V4-Flash-Vision-Exp. Those old names now route to V4.1-Flash. It said third-party tests put the new model ahead of V4-Pro on performance, cost, speed, and total runtime, and that it is phasing V4-Pro out. From 04:00 UTC on September 14, 2026, deepseek-v4-pro requests will run on V4.1-Flash at Flash prices until a V4.1-Pro ships.

New API prices took effect at 04:00 UTC on September 10. Off-peak rates stay at half of peak. The lab did not put the new dollar table in the blog post. Official partners named in the note are WorkBuddy, including CodeBuddy, and OpenCode.

The Hugging Face card adds training detail the blog skips. The model was trained from scratch on a 45 trillion token multimodal corpus, with sparse attention at 64,000 tokens and context stretched to 1 million tokens after 34 trillion. Post-training follows supervised fine-tuning, reinforcement learning, and on-policy distillation. Reasoning effort is an integer from 1 to 100. Each mixture-of-experts layer has 1 shared expert and 384 routed experts, with 6 routed experts active per token.

Decoded Take

DeepSeek is using a cheaper, smaller-activation model to retire its own flagship on the API. That is a pricing decision dressed as a research drop. If V4.1-Flash really sits ahead of V4-Pro for agent work, the 1 million token window and the thin KV cache become the product, not the 552 billion parameter headline. The risk is a quiet downgrade for customers who bought V4-Pro for knowledge work the base-model table still gives to the larger checkpoint. Watch the September 14 reroute and whether V4.1-Pro arrives before those users complain. The MIT weights matter only if the inference stack for a 552 billion parameter encoder-decoder is actually usable outside Hangzhou.