---
id: "lib-2026-10-17-deepseek-v4-1-flash-1017-update-opening-up-the-in"
type: "library"
title: "DeepSeek-V4.1-Flash-1017 Update; Opening up the in-house search API"
lang: "ZH"
org: "DeepSeek"
source_type: "Other"
date: "2026-10-17"
week: "2026-W42"
url: "https://nosov.ai/library/lib-2026-10-17-deepseek-v4-1-flash-1017-update-opening-up-the-in/"
original_url: "https://api-docs.deepseek.com/updates"
tier: "B"
tier_name: "Useful, limited"
tier_reason: "First-party DeepSeek changelog announcing a model update and a public search API; primary source for what changed, but all benchmark scores are the vendor's own with no evaluation detail beyond harness notes."
summary_basis: "full-text"
original_minutes: "3"
tags: ["ai technology", "model release", "coding benchmarks", "search api"]
frames: ["verification-capacity"]
---

# DeepSeek-V4.1-Flash-1017 Update; Opening up the in-house search API

Tier B (Useful, limited): First-party DeepSeek changelog announcing a model update and a public search API; primary source for what changed, but all benchmark scores are the vendor's own with no evaluation detail beyond harness notes.

Summary based on the full text.

## In short

DeepSeek's API changelog entry of 17 October 2026 (published in Chinese and English) covers two things. DeepSeek-V4.1-Flash-1017 replaces V4.1 Flash on the web, app and API at the same endpoint and price, with higher scores on agent and coding benchmarks and shorter outputs. DeepSeek also opened its in-house search engine as a Search and Fetch API.

## Key numbers

- 31.2 to 42.3: Terminal-Bench 4.0 score, V4.1 Flash (10 September) versus V4.1-Flash-1017
- 74.2 to 77.0: DeepSWE v1.1 score, same comparison
- 36.8 to 42.6: HLE without tools, same comparison
- hundreds of millions: search requests a day that DeepSeek says its in-house search engine handles

## What it reports

- Same model name (`deepseek-flash`), same API method and same price; web and app models were updated too.
- Selected scores against the 10 September V4.1 Flash release: Terminal-Bench 4.0 31.2 to 42.3; DeepSWE v1.1 74.2 to 77.0; ProgramBench 20.3 to 29.1; NL2Repo-Bench 65.4 to 68.2; Agents' Last Exam 31.8 to 35.4; HLE without tools 36.8 to 42.6.
- Harness notes: Terminal-Bench 4.0 run with the Claude Code framework, DeepSWE with Mini-SWE-V2; the bracketed HLE scores (44.0 and 39.1) are on the text-only subset.
- Task runtime and output length were reduced, which DeepSeek calls a new performance-cost Pareto optimum. The model was tuned for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy; users are told to upgrade to DeepSeek Harness 0.2.1-rc.1 or later.
- Search: an AI-native general search engine said to index hundreds of billions of web pages, serve hundreds of millions of requests a day, and filter SEO spam, GEO manipulation and low-quality AI-generated content.
- Earlier entries on the same page: V4 Pro API kept running after 14 September 2026; peak and off-peak pricing from 16 August 2026, with off-peak at half price.

## How they got it

Vendor-run benchmarks, reported as single numbers. No sampling settings are given for this release beyond the harness notes, and no independent reproduction.

## Limits

- All scores are self-reported; some benchmarks (Automation-Bench, Apex Bench) have version changes between entries, so not every comparison is like for like.
- Search-engine scale and spam-filtering claims are unverified.
- A changelog, not a technical report: no architecture or training detail.

## Why it matters for companies

A same-price upgrade with large gains on agentic coding benchmarks changes the cost comparison for teams running coding or automation agents on cheaper models, though the numbers need checking on your own tasks. Tuning for specific harnesses is a reminder that model scores now depend on the agent framework around them. The search API adds a non-US option for grounding agents, with data-residency questions to settle first.

## Key passages in translation

> API 调用方式与价格保持不变，仍使用 deepseek-flash 即可访问最新模型。

The API calling method and price remain unchanged; use deepseek-flash to access the latest model.

> 在提升性能的同时，我们还缩短了模型的任务运行时间与输出长度，让用户能够以更低成本、更快速度完成任务

While improving performance, we also shortened the model's task runtime and output length, so users can complete tasks at lower cost and higher speed.

> 针对 SEO 作弊、GEO 操纵和低质 AIGC，我们通过严格的识别机制进行过滤

We filter SEO cheating, GEO manipulation and low-quality AI-generated content through strict detection mechanisms.

Original: https://api-docs.deepseek.com/updates