DeepSeek V4.1 Flash

DeepSeek V4.1 Flash shipped on September 10, 2026 and immediately became the default across the DeepSeek API. It is a 552 billion parameter mixture-of-experts, roughly double the 284 billion in V4 Flash, trained from scratch on 45 trillion multimodal tokens and built on an asymmetric Causal Encoder-Decoder: a 20-layer causal encoder feeding a 20-layer decoder, so only about 8 billion parameters activate per token on input and 16 billion on output. A separate 196 billion parameter Engram memory table sits outside that count. Context is 1,048,576 tokens with up to 384K output and native image input, and the weights are MIT licensed on Hugging Face. DeepSeek retired V4 Flash and V4 Flash Vision Exp on launch day, routing their old model names to V4.1 Flash, and announced that from September 14 every deepseek-v4-pro request would follow. That second half did not happen. On September 11, in response to user demand, DeepSeek withdrew the V4 Pro phase-out: its changelog now says it will continue providing API services for DeepSeek V4 Pro after September 14, 2026 with the billing method unchanged, so a request naming V4 Pro is still answered by V4 Pro at V4 Pro rates. The benchmark claim needs one correction to the launch coverage: V4.1 Flash beats V4 Pro on the agentic rows, not on all of them. DeepSeek reports Terminal-Bench 2.1 at 90.6 against 87.9, DeepSWE v1.1 at 74.2 against 62.7, CyberGym at 88.1 against 83.3, NL2Repo-Bench at 64.0 against 61.5, HLE with tools at 63.9 against 60.0, and a 3,471 Codeforces rating, while V4 Pro keeps GPQA Diamond (92.4 to 90.9) and HLE without tools. Every figure is self-reported. Pricing splits by the clock: $0.30 per million input tokens and $1.20 output at peak, half that off-peak at $0.15 and $0.60, with cache hits at $0.006 and $0.003. Peak is Monday through Friday, 01:00 to 04:00 and 06:00 to 10:00 UTC. Even the peak rate sits below the $0.44 and $1.32 that V4 Flash charged, so the upgrade doubles as a price cut. Budget the peak number if your workload cannot choose its hours, and treat the $0.003 cache figure as a discount on reused context only.