Just now, DeepSeek fully rolled out its new Flash model: all 11 benchmark scores are up, and the cost per task is cut in half. DeepSeek has pushed its new model out to everyone. The official website, App, and API have all switched to DeepSeek-V4.1-Flash-1017, with the weights released open source simultaneously on Hugging Face and ModelScope, still

💡 What You Will Learn

Just now, DeepSeek fully rolled out its new Flash model: all 11 benchmark scores are up, and the cost per task is cut in half. DeepSeek has pushed its new model out to everyone. The official website,

📜 Table of Contents

• Benchmarks all go up, the bill goes down instead
• How the wish-driven programming was trained this time
• The 763B weights are open sourced, but you'll probably only be able to call the API
• Casually selling its own search engine too
• For ordinary users

Just now, DeepSeek has fully rolled out the new Flash model: all 11 benchmarks up, per-task cost cut in half

DeepSeek has pushed the new model out across the board. The official website, App, and API have all switched to DeepSeek-V4.1-Flash-1017, with weights released open source on Hugging Face and ModelScope under the same MIT license. It's only been five weeks since the last version, V4.1-Flash, was released.

Benchmarks all go up, the bill goes down instead

In the official comparison table, all 11 benchmarks are higher than the previous version. Here are a few lined up against the top models:

These are DeepSeek's own self-reported results. Terminal-Bench 4.0 uses the Claude Code framework, DeepSWE uses the mini-swe-v2 framework, at temperature 1.0. A full third-party reproduction hasn't come out yet, so keep that in mind when you look at these numbers.

More interesting than the scores is the cost column. DeepSeek provided a performance-versus-cost chart on DeepSWE v1.1: the new version's max setting hits a 77% pass rate at under $0.2 per task on average; the previous version had a 74.2% pass rate at about $0.5 per task. Higher score, less than half the money.

DeepSeek calls this a "Pareto improvement in performance and cost." The old rule was that if an Agent wanted to do well, it had to think through more rounds and spit out more tokens; the 1017 version cut down both task runtime and output length, saving on token costs and waiting time. The API's unit price and calling method are unchanged, and the model name is still deepseek-flash — the savings come from "saying less."

How the wish-driven programming was trained this time

The four cases officially demoed are all pretty big: a 3D scene of the Forbidden City you can browse in real time and switch viewpoints in, a web-based racing game with a complete flow, a recreation of Portal 2 built on the Godot engine, and a single 3D scene that plays out urban change from prehistoric times to the modern day. The conditions were clearly specified, with very few prompts and reference materials, and it ran straight through in very short interaction rounds inside DeepSeek Harness.

What supports it is self-feedback reinforcement learning. The model doesn't rush to turn in its paper after writing the first version — it uses its own internal confidence and self-evaluation as reward signals, and goes back to fix things when it spots something wrong. The whole pipeline needs neither human evaluators nor an external reward model.

For individual developers, what matters is prototyping cost. Getting an idea running as a prototype you can look at and play with, then deciding whether to keep investing in polishing it — that path used to take days, and now it can start from a single sentence.

The 763B weights are open sourced, but you'll probably only be able to call the API

On the Hugging Face model page, this version's parameter count is listed as 763B, with 48 safetensors shards, FP8 quantization, and a total public weight size of around 510GB. One developer did the math: never mind an RTX 4060 — even a single 4090 or 5090 can't hold the full version, and to keep inference speed and multi-user concurrency, you'd have to move to a multi-GPU server.

The open-source approach hasn't changed, and neither has the barrier. Most people will still actually go through the API. This time DeepSeek even thought of the client for you: eligible accounts get 6 yuan in bonus credit on their first login to DeepSeek Harness, valid for 7 days after it arrives. The official announcement also specifically adapted DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy — Harness needs to be upgraded to 0.2.1-rc.1 or higher to fully take advantage of the new model's optimizations.

Casually selling its own search engine too

Also opened up the same day were the Search and Fetch interfaces, backed by DeepSeek's self-developed search engine. The official word is that it indexes hundreds of billions of web pages globally, handles hundreds of millions of requests per day, and has identification and filtering for SEO cheating, GEO manipulation, and low-quality AIGC.

The pricing is straightforward: both interfaces cost 5 yuan per thousand calls, billed on successful requests — failures, empty results, and timeouts don't count, and returning more results doesn't cost extra, with each account limited to 5 calls per second. For comparison, Perplexity's Search API is tiered by the amount of retrieved content, costing $5 to $14 per thousand calls, making DeepSeek about 85% cheaper. For building retrieval-type applications going forward, one key gets you through retrieval plus generation, and you're left with just one bill.

One date to remember: right now, using the DeepSeek API in DeepSeek Harness, Claude Code, and Codex, the Search and Fetch calls that go through server-side tool invocation aren't billed yet — starting at 0:00 on November 1, they'll be charged per call. If you like having your Agent search the web, this cost will show up on your bill in two weeks.

For ordinary users

The web and App have already automatically switched to the new model — you don't need to do anything. For things like writing code, making small games, and building 3D scenes, you'll feel the speed and savings for yourself. One content creator ran 7 development tests in a row in Harness, burned 619 million tokens, and the bill came to 30.04 yuan.

One thing to watch out for is "shorter output." Concise answers save money for an Agent, but whether that's good for a human depends on the task type. For the work you have on hand, do you want it to say less and just hand over the thing, or do you want it to lay out its reasoning clearly?