DeepSeek-V4.1-Flash-1017

Overview

DeepSeek's 763B MoE update with a native 1M context window - beats GLM-5.3 on Terminal-Bench 4.0 (42.3) and DeepSWE (77.0) while cutting output length and task time. Vision-capable; ships without a Jinja chat template, so use the deepseek-recipe toolkit for prompting.

Good for

• agentic
• chat
• coding

Recommended hardware

≈ 640 GB VRAM to run this 763B model at the FP4-FP8 quantization.

Needs a multi-GPU server node (e.g. 8x 80 GB H100, or 4x 192 GB)

Links

• View on HuggingFace
• Run on Infersec