> ## Content Index
> Fetch the complete content index at: https://globalfeed.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# DeepSeek opens an updated V4.1-Flash that jumps on coding and agent tests
- URL: https://globalfeed.ai/en/deepseek-opens-an-updated-v4-1-flash-that-jumps-on-coding-and-agent-tests/
- Published: 2026-10-17T20:30:10.000Z
- Updated: 2026-10-17T20:30:10.000Z
- Description: DeepSeek has released a new version of V4.1-Flash on Hugging Face under the MIT license. The company calls it a minor update, but coding and agent scores on the model card rose sharply, and its search and fetch APIs are now open to all.
- Author: GlobalFeed Editor
- Tags: DeepSeek, V4.1-Flash, open model, coding, x-deepseek_ai, dil-en, elle

On October 17, 2026, DeepSeek released [DeepSeek-V4.1-Flash-1017](https://api-docs.deepseek.com/news/news261017?ref=globalfeed.ai), a new version of the V4.1-Flash model announced in September. The [model card](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017?ref=globalfeed.ai) calls it a minor update, but the gains on coding and agent tests stand out. The model is available under the MIT license on Hugging Face and ModelScope to run on your own servers, and it is live on DeepSeek's website, app and API.

## Big gains on coding and agent tests

According to the table on the model card, the new version rose from 31.2 to 42.3 on Terminal-Bench 4.0, which measures work in a terminal, from 20.3 to 29.1 on ProgramBench and from 64.6 to 72.1 on Apex Bench. It scores 77.0 on DeepSWE v1.1, which measures fixing software issues; in the same table Claude Opus 5 scores 74.0 and GPT-5.6 Sol 73.0\. On Humanity's Last Exam without tools it rose from 36.8 to 42.6, where Opus 5 still leads with 56.3\. DeepSeek also says the model finishes tasks faster and with shorter outputs, which lowers cost.

## What the model offers

The model has 763 billion parameters in total, with a recommended context of 1 million tokens. It reads images as well as text and supports tool calling, a thinking mode and an adjustable reasoning level. It runs with vLLM, SGLang and Docker Model Runner, and DeepSeek also shared a Rust-based toolkit for production use. The model was trained with reinforcement learning on its own feedback for end-to-end project building; the announcement shows a 3D model of the Forbidden City, a web racing game and a Portal 2 remake in Godot. Running a model this size locally still takes very powerful servers.

## Harness and search APIs

The model is tuned for DeepSeek Harness, the company's agent framework, and for Tencent's CodeBuddy and WorkBuddy. On the same day, DeepSeek opened its Search and Fetch APIs, built on its own search engine, to all users. These APIs are paid; server-side tool calls inside DeepSeek Harness, Claude Code and Codex are free for DeepSeek API users until October 31\. Current prices are on DeepSeek's page.

The table's results are figures shared by DeepSeek, and some tests were run with a specific setup (Claude Code for Terminal-Bench, mini-swe for DeepSWE). The model card says nothing about language support. It is worth trying the model on your own codebase and watching for comparisons from independent researchers.