V4.1 Flash 1017 规格解读：552B MoE、1M 上下文、KV 占用降至 V4-Flash 四分之一
官方信息梳理：1017 是 Pareto 更新，全 benchmark 超 GLM 5.3 与 Kimi K3；552B 多模态 MoE、1M 上下文、推理强度 1–100 可调，KV 约 890 字节/token，建议 Harness v0.2.1-rc.1+。
DeepSeek describes 1017 as a Pareto update: stronger benchmarks, faster task completion, and shorter outputs, with reinforcement learning from self-feedback aimed at complete project generation. Surpassing GLM 5.3 + Kimi K3 on every benchmark 👀 It is live in the app/API and as downloadable weights; @deepseek_ai specifically recommends Harness v0.2.1-rc.1 or newer, which matters because agent scores are increasingly harness-dependent rather than model-only properties. The underlying V4.1-Flash architecture is a 552B-backbone multimodal MoE with 8B parameters active in prefill, 16B in decode, a 1M-token context, and continuously adjustable reasoning effort from 1–100. Its Causal Encoder-Decoder design plus bounded SWA replay and FP4 KV caching reduces persistent KV footprint to about 890 bytes/token—roughly one quarter of V4-Flash—making long-context agent workloads the more interesting claim than headline parameter count. Go download & install the new 1017 update Do you think it’s benchmaxed or will it hold up against open-sourced models with its intelligence? 🧐
这条是引用着另一条写的：x.com/deepseek_ai/status/2111375640717033777