DeepSeek ships a new open AI update: faster, and it costs cents

Chinese company DeepSeek announced on October 17, 2026 a new release of its AI model, DeepSeek V4.1-Flash-1017. It is an update to the Flash version released on September 10, 2026 — not a brand-new architecture, but the company says it is substantially improved.

The biggest news is that the weights are open. DeepSeek posted them on Hugging Face under the MIT license, which lets anyone use, modify, and even commercialize them for free. According to the model card, the weights total 510 GB and include image understanding alongside text.

The company claims the release scores better on all 11 benchmarks it reports, per its own table. For example, on the DeepSWE v1.1 coding test it scores 77.0, up from 74.2, and on Terminal-Bench 4.0 it scores 42.3, up from 31.2. In DeepSeek's own comparison, it beats Kimi K3 on 10 of 11 tests and GPT-5.6 Sol on 9 of 11 — but still trails Claude Opus 5 on 7 of 11. One important caveat: these are DeepSeek's own numbers, not independently verified. However, because the weights are public, outside researchers can rerun the tests themselves.

DeepSeek also says the release finishes tasks faster and produces shorter outputs — which cuts costs further. Pricing is unchanged: $0.15 per million input tokens (off-peak) and $0.60 per million output tokens. The Decoder's write-up shows one DeepSWE task costing just under $0.20 on the new release — compared with around $0.50 for the previous one — while rivals range from $2 to over $10 for the same task.

Alongside the model, DeepSeek opened two new services to developers: the Search and Fetch APIs. Search returns ranked web results, and Fetch returns a single page as markdown. They cost $0.75 per 1,000 calls, but are free for now through server-side tools in DeepSeek Harness, Claude Code and Codex until October 31, 2026.

But open weights don't mean everyone can run it at home. 510 GB needs multi-GPU server hardware — not an ordinary laptop. And until independent tests appear, DeepSeek's figures remain company claims, not a final verdict.

What does this mean for you?

For small businesses: Local developers can now build AI-powered tools at lower cost than before. The MIT license means Swahili voice assistants (like Oria) or customer-service systems can run for cents per task — which could make them cheaper for you.

For students: The DeepSeek app is free online and on phones, and this release is claimed to be better at coding and task-solving. It's a low-cost option for learning to code or tackling assignments — but remember: AI can still get things wrong, so verify key facts.

For software developers: You can download the weights under the MIT license and build a commercial product on top without paying licensing fees. DeepSeek says it tuned this release for coding agents (DeepSeek Harness, Tencent CodeBuddy and WorkBuddy).

A caution: The benchmark figures are DeepSeek's own and unverified by outsiders. Don't swap out your critical model immediately — test first and wait for independent evaluations.