DeepSeek V4.1 Flash 1017

DeepSeek V4.1 Flash 1017 is a minor update to DeepSeek V4.1 Flash, keeping the same natively multimodal mixture-of-experts architecture — 763B total parameters with 16B active, a 1M-token context, and an FP4 KV cache of under 1KB per token — while sharpening reasoning and agentic behaviour. Across DeepSeek's published benchmarks it improves on V4.1 Flash on every task reported: HLE without tools 42.6 vs 36.8, Terminal-Bench 4.0 42.3 vs 31.2, DeepSWE v1.1 77.0 vs 74.2, NL2Repo-Bench 68.2 vs 65.4 and Apex Bench v1.1 72.1 vs 64.6, with vision results of 83.0 on Chartography and 92.5 on BabyVision when tools are enabled. It also cuts task execution time and output length, lowering cost per completed task and making it the better default for long-horizon agent runs, repository-scale engineering, and document-heavy multimodal work. Weights and reference encoding are MIT-licensed; the release ships without a Jinja chat template, providing a Python reference encoder and the deepseek-recipe toolkit instead. Recommended sampling is temperature 1.0, top_p 1.0 and at least 256K max tokens.