DeepSeek-V4.1-Flash: Smarter, Faster, More Efficient

Summary

DeepSeek's announcement of V4.1-Flash, a natively multimodal open-weight model that, by DeepSeek's own benchmarks, beats its larger flagship V4-Pro. It drops the decoder-only design for a "causal encoder-decoder" mixture-of-experts architecture and cuts KV-cache memory again. Weights are on Hugging Face (model card) under MIT.

Key points

• 552B parameters in total. It activates 8B parameters per token when reading input (prefill) and 16B when generating (decode).
• Architecture: a 20-layer causal encoder followed by a 20-layer decoder, both MoE. Uses Compressed Sparse Attention 2 (CSA2) and an FP4 KV cache.
• KV cache uses about a quarter of the HBM and an eighth of the SSD storage of V4-Flash.
• Pre-trained on about 45T multimodal tokens. It has its own vision encoder (DeepSeek-ViT) and accepts images and text.
• Reasoning effort is adjustable on a 1–100 scale. Speculative decoding uses "DSpark".
• 1M-token context. In the API it is the deepseek-flash model, and since 2026-09-14 deepseek-v4-pro requests are routed to it. Prices were cut from 2026-09-10.

Why it matters

The largest open-weight lab has moved its newest model away from decoder-only transformers, and a mid-sized model now beats its 1.6T flagship. The open-weight frontier is no longer simply the largest model.