DeepSeek V4.1 Flash 1017 ships for local and gateway agent workloads

What changed · following since Oct 18 Oct 25 — vLLM demonstrates that a general-purpose engine can serve DeepSeek V4.1 Flash at scale with throughput within 6% of DeepSeek's own DeepInfer engine on GB300 NVL72 hardware.

DeepSeek V4.1 Flash 1017 is available both as a locally deployable model artifact and through Vercel AI Gateway, lowering the friction between evaluation and production use. The update is positioned around stronger reasoning, coding, agentic automation and image understanding, while producing shorter outputs and completing tasks faster. Independent benchmarking now shows the model scaling efficiently on GB300 NVL72 hardware with vLLM, achieving throughput within 6% of DeepSeek's own serving engine. For you, the practical change is a model option that can be tested locally, operated through a gateway with controls, and served at scale on commodity inference infrastructure.

• The model card lists DeepSeek V4.1 Flash 1017 as an image-text-to-text Transformers model with Safetensors files, an MIT license and 8-bit fp8 precision; it documents Transformers, vLLM and Docker usage paths.
• The update is described as improving terminal work, whole-repository coding and multi-step automation, with more accurate chart reading and visual reasoning when tools are available.
• Vercel AI Gateway exposes the model as `deepseek/deepseek-v4.1-flash-1017` through one API key, with image input, request traces and spend tracking.
• AI Gateway adds operational controls around the model, including retries, failover, performance routing, budgets, custom reporting and Zero Data Retention support.
• On one NVIDIA GB300 NVL72 rack, vLLM with P/D disaggregation and Wide-EP serves the model at 954K tokens/s per GPU at a P90 of 196 tokens/s per user, with decode throughput within 6% of DeepInfer across batch sizes 8–256.
• A specialized decode path reaches 826 tokens/s for a single user, and the setup costs $38.96 per GPU-hour at DeepSeek's off-peak prices.

Bottom line If you run coding or agent workflows, test this model against representative tasks locally, then compare latency, output tokens and failure recovery before routing production traffic through it.