DeepSeek updates V4.1 Flash and opens its search API

The model got a little better. The search API is the part you should try this week.

DeepSeek posted DeepSeek-V4.1-Flash-1017 on Hugging Face under the MIT license, and its API release notes date the update Oct. 17. The company calls it a minor update to V4.1-Flash. On the same day it opened its in-house search engine to developers as a paid API.

A small update, scored by DeepSeek

The Hugging Face page lists 763B parameters and a recommended 1M-token context window. It is an open-weight model: you can download it, run it and modify it under MIT terms.

DeepSeek reports these gains over the previous V4.1-Flash, on its own runs:

• DeepSWE v1.1: 77.0, up from 74.2
• Terminal-Bench 4.0: 42.3, up from 31.2
• ProgramBench: 29.1, up from 20.3

On DeepSeek's own table, the new build beats GPT-5.6 Sol on DeepSWE (73.0) and Terminal-Bench 4.0 (39.9). Opus 5 still scores higher on Terminal-Bench 4.0 (51.8), ProgramBench (37.0) and HLE without tools (56.3). These are the company's figures, not independent results.

DeepSeek also says the update cuts task runtime and output length, so tasks cost less to complete. API pricing and the calling method are unchanged, and you reach the model as deepseek-flash. The Hugging Face page says no third-party inference provider hosts the weights yet.

Search is the new thing to build on

DeepSeek's Search and Fetch guide describes two endpoints. One searches the web. The other fetches a page. In its release notes, DeepSeek says its index covers hundreds of billions of web pages and filters SEO spam and low-quality AI-generated content.

The price, per endpoint:

$0.75 per 1,000 successful requests.

Failed calls, timeouts and empty results are not billed. Each endpoint is capped at 5 queries per second per account. Server-side tool calls made through the DeepSeek API from DeepSeek Harness, Claude Code or Codex are free until 00:00 Beijing time on Nov. 1, when per-request billing starts.

What it means if you run agents

Yes, another Chinese open model with a benchmark table. The reason to care is cost.

TopFive reported yesterday that open-weight models carried seven in ten tokens on Vercel's AI Gateway. A model that finishes agent tasks in fewer tokens at the same price pushes that further. A search API at $0.75 per thousand calls gives an agent a web tool without a second vendor.

Two cautions.

First, search is a new dependency on a provider based in China. Check where your queries go before you send customer data through it.

Second, the benchmark gains came on harnesses DeepSeek chose: Claude Code for Terminal-Bench 4.0 and Mini-SWE-V2 for DeepSWE. Your harness may tell a different story.

If you already route cheap tasks to DeepSeek, nothing breaks. Same model name, same price. Run your own eval set against the new build before you trust the jump.

What to watch

Nov. 1 is the date. That's when free search calls start costing money, and when you'll learn whether developers keep using it once they pay.

Watch for independent Terminal-Bench runs too.

Test it before the meter starts.