DeepSeek V4.1 Flash 1017 552B vs DeepSeek V4.1 Flash 552B

Analysis

DeepSeek V4.1 Flash 1017 is a minor update to the September 10 V4.1 Flash, released five weeks later on October 17, 2026. Nothing about the hardware bill changed: the configuration file is identical and the weights are the same 510.3 GB, so this comparison is not about VRAM at all. It is about what five weeks of extra post-training bought, and about the one thing the new build lacks so far, which is any independent check.

Everything structural is the same. Both are 552B-parameter mixture-of-experts models with 16B active per token on decode and 8B on prefill, 40 layers split into a 20-layer encoder and a 20-layer decoder, a 1M-token context window, native image input and an MIT license, and both hold a full-window KV cache to the same 2.3 GB on this site's math, so at any quantization or context length their VRAM totals are identical (about 571.6 GB at an 8k context for the native release). What differs is the post-training, and DeepSeek published a side-by-side table with the release. The new build is ahead on every row the two share, led by Terminal-Bench 4.0 (31.2 to 42.3), ProgramBench (20.3 to 29.1) and Humanity's Last Exam without tools (36.8 to 42.6), while DeepSWE v1.1 moves only from 74.2 to 77.0. DeepSeek also says the update finishes tasks faster with shorter outputs, and it tuned the build for its own agent harness and for Tencent's CodeBuddy and WorkBuddy. Two gaps limit how far to trust the comparison. Every one of these figures is DeepSeek's own run, and no independent lab had re-scored the 1017 build when this page was written. And any GPQA Diamond figure shown for the two builds is the same 90.9 only because this site carries the September score over to 1017 so they rank level; DeepSeek's 1017 card has no GPQA row. Its card also reports Terminal-Bench 4.0, a newer suite than the Terminal-Bench 2.1 (90.6) behind the September build's entry, so there is no like-for-like Terminal-Bench figure to compare. Availability differs slightly too. Community conversions of the September weights, such as a GGUF of about 169 GB and an MLX build of about 239 GB, are weeks old, while conversions of the 1017 weights (an NVFP4 repository of 429.4 GB and an MLX build of 432.0 GB) appeared within hours of release and are unverified, and the two sets are not interchangeable.

Bottom line: If you call DeepSeek's API through the deepseek-flash name you are already on the 1017 build, and DeepSeek's own table says that is a gain for agent and terminal work, with the caveat that it is DeepSeek's table. If you host the weights on a datacenter node, pull the 1017 checkpoint for agent work: it costs exactly the same memory as the September one. If you were hoping for a quantized build that fits a smaller machine, neither has a verified one, and the September community conversions are the more mature but cannot be loaded as the new weights. Stay on the September build only if you need its published GPQA Diamond or Terminal-Bench 2.1 numbers for a like-for-like record.