Kimi vs DeepSeek: Kimi scores higher on the Intelligence Index, DeepSeek costs 7x less per task and won our 30 language test

Which is better, Kimi or DeepSeek?

Kimi K3's five point lead costs 7.4 times as much per task. Artificial Analysis runs each model through the same ten tests, from Terminal-Bench 4.0 to Humanity's Last Exam, and prices every task at the maker's list rates. On Humanity's Last Exam, one of those ten, Kimi K3 scores 47% to the 39% of DeepSeek V4.1 Flash, read on 12 October 2026.

You also wait longer for Kimi. It took Kimi K3 1,111 seconds to finish an average task, which is about 18 and a half minutes. It took DeepSeek's Flash 304. Moonshot's API wrote 41 tokens a second, and DeepSeek's wrote 205.

All of the scores, costs, times and speeds in the table are from Artificial Analysis (Intelligence Index v4.3.2). The prices are what each maker charges on its own API at full rate: Moonshot's price list for Kimi and DeepSeek's for DeepSeek. If you came here for Kimi K2 vs DeepSeek, the table has that answer for you as well. Kimi K2.6 and K2.7 Code score 27 and 26, so both trail DeepSeek V4 Pro 0813 on score, and both cost more per task than the Flash.

Is Kimi or DeepSeek better in other languages?

DeepSeek is, and by a wide margin on our own test. Whizi's Multilingual LLM Bench asks 150 questions that we wrote from scratch, five in each of 30 languages, and grades every answer against a sourced key. We ran each model once at low reasoning and finished on 2 October 2026. DeepSeek V4 Pro scored 44 out of 100 and V4.1 Flash scored 42. Kimi K3 scored 31.3.

That's the reverse of the Intelligence Index order. Kimi K3 trails DeepSeek V4 Pro by 12.7 points, and the margin of error at this sample size is 7 to 8 points for each model. Most of the gap is in idioms and in grammar, script and calendar questions. On local facts, Kimi K3 and DeepSeek's Flash came within two points of each other.

DeepSeek won with a handicap. Every answer got a budget of 8,000 tokens, and the two DeepSeek models used all of it on thinking for 22 of the questions between them. Each of those scored zero. Kimi K3 ran out on none.

Five questions per language is too few to rank one language alone, so read the next part as a hint. Kimi K3 beat DeepSeek V4.1 Flash in 6 of the 30 languages, Arabic and German among them, and lost in 16. Both labs are Chinese, and in Simplified Chinese the Flash got all five questions right. Kimi K3 got two.

Is Kimi or DeepSeek better for coding?

DeepSeek V4.1 Flash is, if your agent works in a terminal. On Terminal-Bench 4.0, one of the ten tests in the index, it scores 27% to the 13% of Kimi K3. It also leads AutomationBench-AA, 69% to 58%. Kimi K3 wins SciCode, which asks for scientific code from a written spec, by 59% to 52%.

Moonshot's coding model is the surprise. Kimi K2.7 Code scores 1% on Terminal-Bench 4.0 and 48% on SciCode, so on that test it trails both DeepSeek V4 models. Speed counts for more in a coding loop than in a chat, too. At 205 tokens a second, DeepSeek's Flash writes five times as fast as Kimi K3. Claude vs DeepSeek runs the same tests against a closed model.

How much do Kimi and DeepSeek cost?

DeepSeek V4.1 Flash costs a tenth as much per token, or less, and V4 Pro 0813 under half. As of October 2026, Moonshot's price list charges $3.00 for each million tokens you send Kimi K3 and $15.00 for each million it writes. DeepSeek charges $0.30 and $1.20 for V4.1 Flash, and $1.32 and $3.96 for V4 Pro 0813. Both of them cut the price of input that's in the cache, to $0.30 a million on Kimi K3 and to $0.006 on the Flash.

DeepSeek also halves every rate off peak. Its peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, so a job you send on a weekend pays $0.15 and $0.60 on the Flash. Moonshot's page doesn't list an off peak rate, and neither page gives free tokens for these models.

The apps split the same way. On the US App Store, the DeepSeek app is free and lists no in-app purchases. The Kimi app is free to download and sells memberships inside the app, listed from $19.00 up to a $374.99 annual pass, read in October 2026.

Can you build a product on Kimi or DeepSeek weights?

Yes, on both, but only DeepSeek's weights come with nothing to sign. DeepSeek ships V4.1 Flash and V4 Pro 0813 under the plain MIT license on Hugging Face. Kimi K3 comes under its own Kimi K3 License, which starts from MIT wording and adds two conditions. We read both files on 12 October 2026.

The display rule is the one a growing app hits first. Past either line, the license says the name "must be prominently displayed on the user interface of such product or service." Two uses are exempt from both rules: internal use where no outsider gets the model or what it writes, and use through Moonshot's own products or its certified inference partners.

Size matters too, if you plan to host the weights yourself. Kimi K3 has 2.8 trillion parameters with 104 billion active per token. DeepSeek V4.1 Flash has 552 billion with 16 billion active, so it's about a fifth the size. The open source LLMs guide covers what it takes to run weights like these.

Where do Kimi and DeepSeek keep your data?

Both apps keep it in China, and Kimi's API keeps it in Singapore. The Kimi privacy policy for kimi.com and the Kimi apps, updated 24 August 2026, names Beijing Moonshot Technology and says: "We store your personal information within the territory of the People's Republic of China." The DeepSeek privacy policy says it will "collect, process and store your Personal Data in People's Republic of China."

Moonshot runs its developer platform through a different company. That platform's policy names MOONSHOT AI PTE. LTD. and says it keeps what it collects on servers in Singapore. DeepSeek doesn't publish a separate policy for its API, so China is the only location it gives.

Neither app policy describes a switch in the settings for training. Kimi tells you to contact customer service, and DeepSeek lists an opt out you use by email. If you run the open weights on a host you trust, your data goes to neither company. Is DeepSeek safe covers the rest of DeepSeek's policy.

Should you pick Kimi or DeepSeek?

Pick Kimi K3 if the top reasoning score is the thing you need. Pick DeepSeek if you care more about the cost of a task, about speed, or about a language other than English.

Whizi has both of them on its Pro plan, which is $29.99 a month with 2,500 credits. Kimi K3 costs 10 credits a message there and DeepSeek V4.1 Flash costs 1, and the side by side view sends one prompt to both. So Pro buys 250 Kimi K3 messages, or 2,500 from the Flash. DeepSeek V4 Pro 0813 is on Powerhouse at $49.99 (2 credits), and Starter includes neither family. If DeepSeek is all you want, its own app is free and Whizi isn't. The Kimi and DeepSeek docs list every row.