On October 17, 2026, Hangzhou DeepSeek released DeepSeek-V4.1-Flash-1017, an update to the V4.1-Flash model it had launched only about five weeks earlier, on September 10.

The company called it a "minor update" that reduces task execution time and output length, yet within a day it had knocked Kimi K3 (Max) off the top of the open-model rankings.

According to Crypto Briefing, arena.ai's October 18 update showed the new build scoring 1670 points on Code Arena WebDev — #1 among open models and #12 overall against closed competitors too — at roughly 10x lower cost than the model it displaced.

The update beat the September version on every benchmark DeepSeek published. DeepSWE v1.1 rose to 77.0 from 74.2, Terminal-Bench 4.0 jumped to 42.3 from 31.2, and Automation-Bench climbed to 60.5 from 54.8, per Crypto Briefing.

Silicon Republic reports DeepSeek recorded 77 on DeepSWE, ahead of OpenAI's GPT-5.6 Sol at 73 and Anthropic's Claude Opus 5 at 75, and says the model shows "substantial improvements" across reasoning, coding, agentic tasks and visual benchmarks.

The caveats matter: SiliconANGLE notes Anthropic's model still led on seven of the 11 benchmarks DeepSeek published, with a nearly 14-point lead on Humanity's Last Exam.

DeepSeek's release framing touts a Pareto improvement in performance and cost: better scores and faster task completion at unchanged prices.

Crypto Briefing reports the API costs $0.30 per million input tokens and $1.20 per million output tokens at peak, and SiliconANGLE notes off-peak pricing at the September launch started at 15 cents per million uncached input tokens.

GIGAZINE reports peak-hour fees of $0.30 input, $0.006 cached input and $1.20 output per million tokens, halved off-peak, and says the model surpasses GPT-6 Astra in coding benchmarks while keeping costs very low.

The update went live through the existing deepseek-flash model ID, so existing customers got the upgrade automatically.

The build is a Mixture-of-Experts design with a 552-billion-parameter backbone plus an additional 196 billion Engram parameters, per Crypto Briefing, which details sparse activation: 8 billion parameters fire during prefill and 16 billion during decoding.

GIGAZINE puts total parameters at 763 billion and warns that is far too large for a personal computer, though Ollama and LM Studio quickly began US-hosted serving with zero data retention.

The model features a Causal Encoder-Decoder structure, native vision, KV cache optimizations and a 1-million-token context window. Weights are published on Hugging Face under the MIT license.

DeepSeek tuned the update for agent software: its own DeepSeek Harness tool, Tencent's CodeBuddy coding assistant and WorkBuddy, the desktop agent for office work, SiliconANGLE reports. The change log recommends Harness version 0.2.1-rc.1 or later.

DeepSeek also opened its in-house search engine to outside developers through a Search API returning up to 50 ranked web results per query, alongside a Fetch API that converts a single web page to Markdown; both cost 75 cents per 1,000 successful requests, with calls through DeepSeek Harness, Anthropic's Claude Code or OpenAI's Codex free until November 1.

The releases follow a nearly five-hour API outage on October 15 that Yicai Global counted among DeepSeek's worst this year, after which the company apologized, credited users and promised to remove single points of dependency.

The release landed days after Meta unveiled Muse Spark Open, an 816-billion-parameter model it called the most capable US open-weight model to date, and after Mistral announced Mistral Large 4, a 1-trillion-parameter multimodal model it says outperforms any open-weight model developed in the US or Europe, Silicon Republic reports.

DeepSeek's repeated price moves keep reshaping that race — its September V4.1-Flash launch sent shares of Chinese rivals MiniMax and Alibaba down in Hong Kong, per Bloomberg as reported by SiliconANGLE.

Eleven models still sit above V4.1-Flash-1017 overall on Code Arena, but a #12 ranking at $0.30 per million input tokens gives engineering teams a strong case for routing coding workloads its way.