This write-up frames V4.1 Flash as a generational replacement rather than a point update, comparing it against the now-retired V4 Flash. The backbone grows from 284B to 552B parameters, while active parameters per token drop from 13B to 8B for input reading and 16B for output generation. The architecture shifts from a Benchmark deltas highlighted include Terminal-Bench 2.1 moving from 82.7 to 90.6, Terminal-Bench 4.0 from 7.0 to 31.2, and DeepSWE v1.1 from 54.4 to 74.2, with context window holding at 1M tokens and license remaining MIT. The piece notes that routing `deepseek-v4-pro` to V4.1 Flash from 2026-09-14 (04:00 UTC) brings a

DeepSeek V4.1 Flash is described as a 552B-parameter Mixture-of-Experts model that activates only 8B parameters per token during prefill and 16B during decoding, supporting a 1M-token context window with up to 384K output tokens and processing images and text natively. Weights are released under the MIT license, enabli The piece reports that V4.1 Flash was benchmarked against OpenAI's GPT-5.6 Sol and Anthropic's Claude Opus-5.0 on four of the hardest agentic benchmarks, scoring DeepSWE v1.1 74.2, AutomationBench 54.8, Agent's Last Exam 31.8, and CyberGym 88.1. Peak input pricing is cited at $0.30 per million tokens. DeepSeek is so co

DeepSeek officially released DeepSeek-V4.1-Flash on 2026-09-10, describing it as the smallest model in a new architecture family with native multimodal visual understanding, designed for higher capability ceiling, faster inference, higher throughput, and scaling to larger models. Published benchmark scores include GPQA On the API side, the model name to call the latest V4.1 Flash is `deepseek-flash`, with native multimodal support. The previous-generation models V4 Flash and V4 Flash Vision Exp have been retired; the legacy names `deepseek-v4-flash` and `deepseek-v4-flash-vision-exp` are temporarily routed to V4.1 Flash. After 12:00