DeepSeek's V4.1-Flash-1017 update scores higher and cuts the cost per task

The new model replaces DeepSeek V4.1 Flash everywhere at once. It now runs the DeepSeek website, the app and the API, and the company also posted the weights for download on Hugging Face and ModelScope.

The license is MIT, one of the most permissive there is. You can download the model, run it on your own hardware and build products on it.

The name follows a date pattern. "1017" means October 17, the release day. Startup Fortune notes that DeepSeek has used the same style before, with snapshots like V4-Flash-0731 and V4-Pro-0813.

DeepSeek first launched V4.1 Flash on September 10. Its change log calls it the smallest model in the company's new architecture family, with native image understanding built in. The 1017 snapshot is a tune-up of that model, not a new generation.

• Context window: 1 million tokens, per the Hugging Face model card
• Maximum output: at least 256,000 tokens
• Inputs: text and images
• Recommended settings: temperature 1.0, top_p 1.0

If you missed the background, we covered how DeepSeek is adapting its software to Huawei's Ascend chips last month.

DeepSeek says the update improves reasoning, coding, agent tasks and visual tests. Its published table shows gains on every benchmark it lists. Here are the main ones, old version against new.

MindStudio lists more of the table. Apex Bench v1.1 rose from 64.6 to 72.1, Automation-Bench from 54.8 to 60.5, Chartography from 78.9 to 83.0 and ZeroBench from 49.0 to 51.0.

The fine print matters. DeepSeek tested Humanity's Last Exam only on its text questions. It ran Terminal-Bench inside the Claude Code framework and DeepSWE inside the Mini-SWE-V2 framework.

The catch: these are DeepSeek's own numbers. The Decoder points out that independent benchmarks haven't come out yet, and that developers often judge code quality differently than a benchmark does.

On some tests, according to DeepSeek. The Decoder reports that in the company's benchmarks the update beats larger rivals, including Claude Opus 5, GPT-5.6 Sol, Kimi K3 and GLM 5.3, in several coding and agent categories.

MindStudio adds the other side. Opus 5 still leads on reasoning tasks. So the honest summary is "a small model that is very good at coding jobs," not "a small model that is better at everything."

Even the previous version was strong at coding. Startup Fortune reports that V4.1 Flash had a Codeforces rating of 3471, ahead of the older V4 Flash at 3289 and V4 Pro at 3348. On DeepSWE it scored 74.2%, edging out Opus 5 at 74.0%.

Because AI costs money per token, and a token is roughly a word or a piece of one. An agent that rambles pays for every extra word. An agent that gets to the point finishes the same job for less.

Think of a taxi that knows the shortcuts. The fare per mile stays the same, but you pay less because the ride is shorter.

DeepSeek kept the API price the same, and the shorter outputs cut the cost of a typical coding task by more than half. The Decoder describes a DeepSeek chart for the DeepSWE coding test:

• New model, top setting: 77.0% at just under $0.20 per task
• New model, low setting: about 72% for about $0.10 per task
• Previous version: 74.2% at about $0.50 per task
• Rivals in the chart: GPT-6 Astra, Gemini 3.8 Flash and Claude Opus 5 at between $2 and more than $10 per task

DeepSeek calls this a "Pareto improvement," a fancy way of saying it got better on one measure without getting worse on another. In plain words: higher scores, lower bills.

The Decoder reports that DeepSeek used a method called reinforcement learning with self-feedback, or RLSF. The model scores its own work, through its confidence or its own critiques, and learns from that signal. No human graders or separate reward model are needed.

DeepSeek says this helps the model build whole software projects from scratch. The examples range from game programming to 3D modeling and design work. The Decoder adds that it can animate short explainer videos.

A harness is the program that wraps a model and lets it act: run commands, edit files, call tools. DeepSeek says it tuned the update for its own DeepSeek Harness, plus Tencent's CodeBuddy and WorkBuddy. It recommends Harness version 0.2.1-rc.1 or later.

Startup Fortune reads this as DeepSeek building its own full stack. DeepSeek Harness, known by its command name dsh, launched on August 13 under an MIT license. The outlet reports it passed roughly 197,000 GitHub stars within eight days.

Tencent's tools aren't exclusive to DeepSeek, though. Citing 36Kr, Startup Fortune says WorkBuddy lets users switch among models like Hunyuan, DeepSeek, GLM, Kimi and MiniMax. CodeBuddy also runs Claude, GPT, Gemini and Hunyuan.

In real life If you already use DeepSeek through Claude Code or Codex, you don't have to change anything. Requests to deepseek-flash now go to the new model at the same price, and the built-in web search is free until the end of October.

DeepSeek is also opening its in-house search engine to outside developers through two new APIs, Search and Fetch. The company says the engine already handles hundreds of millions of requests a day and indexes hundreds of billions of web pages in many languages.

DeepSeek says it filters out SEO spam, attempts to game AI answers and low-quality AI-generated pages. Here is how the pricing works:

• Search: $0.75 per 1,000 successful requests
• Fetch: $0.75 per 1,000 successful requests
• Free period: server-side search in DeepSeek Harness, Claude Code and Codex costs nothing until 00:00 Beijing time on November 1, 2026

DeepSeek's pricing page lists the model as deepseek-flash. The legacy names still work and get routed to the new model at the Flash price.

By the numbers: off-peak rates are half the peak rates. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday. Weekends and Chinese public holidays are off-peak all day.

The big picture: the release lands at an awkward moment for open models from China. The Decoder reports that the Chinese government plans to tighten rules on publishing open-weight models, possibly depending on whether a model ranks among the 10 most capable.

• If you build with AI: coding agents on deepseek-flash should get faster and cheaper per task without any code change.
• If you run models yourself: the MIT license lets you download the weights and use them commercially.
• If you compare models: treat the benchmark wins as DeepSeek's claims until independent tests arrive.
• If you need web search in an agent: the new APIs are free inside Claude Code, Codex and DeepSeek Harness until October 31.