Kimi K3 Review: We Tested It vs DeepSeek V4 Pro and V4.1 Flash

Kimi K3 is Moonshot AI's flagship model, released on July 16, 2026: 2.8 trillion parameters, open weights, image input and a 1M-token context. We gave it the same five tasks we ran on DeepSeek's two models, and it got every one right.

The catch is speed and price. At its default settings Kimi K3 was 2 to 4.4 times slower than DeepSeek V4.1 Flash on every task, and 12 to 32 times more expensive at list prices.

Short answer

• Kimi K3 is accurate: it passed all five tasks, including 30 of 30 coding tests and a search through a quarter-million-token log.
• It always thinks, at maximum effort by default, so it is neither fast nor cheap out of the box.
• DeepSeek V4.1 Flash matched it on every task in a fraction of the time and money.
• Our pick: DeepSeek V4.1 Flash for everyday API work; Kimi K3 when you want an open-weight frontier model or its strengths in browsing and tool-heavy agents.

All three models are on OmniaKey: Kimi K3, DeepSeek V4.1 Flash and DeepSeek V4 Pro.

What is Kimi K3?

It is Moonshot's biggest model so far, and its weights are open. Moonshot's launch post and API docs give these facts (checked October 18, 2026):

Moonshot is candid about where it stands: it says K3's overall performance "still trails the most powerful proprietary models, Claude Fable 5 and GPT 5.6 Sol".

How did it do in our tests?

It got all five right. These are the same tasks we used in our DeepSeek V4.1 Flash 1017 review:

• Parser: write an arithmetic evaluate() with Python's exact rules, scored by our own 30 test cases.
• Invoice: pull 10 fields from a German-English email with European number formats.
• Tool calls: weather and local time for two cities; the right answer is four calls in one reply.
• Long log: find three current badge codes in a quarter-million-token log, skipping an outdated one.
• Chart: read the highest, lowest, average and one difference from an image.

How we ran it: Kimi K3 on October 18, 2026 and DeepSeek's models on October 17, one run per task, default settings, through OmniaKey's OpenAI-compatible endpoint. Cost is each reply's token count times the maker's list price (DeepSeek's off-peak rate). OmniaKey sells API access to all three, so we ran the tests through it; every price here is the maker's, not ours. One run per task, a day apart, is a spot check, not a benchmark.

Against V4 Pro, Kimi K3 wrote the parser 7 times faster (176 seconds against 20.7 minutes) at about the same cost. On the other three shared tasks V4 Pro was cheaper, and faster on two of them.

Why is Kimi K3 slower and pricier?

Two reasons. K3 always thinks, and its API runs at maximum effort unless you ask for less; Moonshot's API takes reasoning_effort as low, high or max. Its list price is also 20 times DeepSeek V4.1 Flash's for input and 25 times for output.

Prices are from Moonshot's pricing page and DeepSeek's (off-peak), checked October 18, 2026. Moonshot says its API hits the cache on over 90% of input in coding work, which cuts long coding sessions a lot. Our tests were single calls, so they had no cache hits.

Kimi K3 or DeepSeek: which should you use?

DeepSeek V4.1 Flash for most work, Kimi K3 for a few specific jobs.

What do Moonshot's own benchmarks say?

K3 leads Claude Fable 5 and GPT-5.6 Sol on several agent tests in Moonshot's model card: BrowseComp (91.2, against 88.0 and 90.4), MCPMark (94.5, against 87.4 and 92.9) and SWE-Marathon (42.0, against 35.0 and 39.0). It trails Fable 5 on FrontierSWE (81.2 against 86.6), GDPval-AA (1686 against 1747) and HLE without tools (43.5 against 53.3).

Read vendor tables carefully. Moonshot's card gives K3 77.8 on ProgramBench, while DeepSeek's own table gives K3 17.5 on a test of the same name. Your own tasks, run the same way on each model, are the better guide.

How do you call Kimi K3?

from openai import OpenAI

client = OpenAI(api_key="YOUR_OMNIAKEY_KEY", base_url="https://api.omniakey.com/v1")

stream = client.chat.completions.create(
    model="kimi-k3",
    messages=[{"role": "user", "content": "Write a Python function that parses ISO 8601 durations."}],
    max_tokens=32000,                 # K3 always thinks: leave room for long answers
    stream=True,                      # long thinking runs can outlast proxy timeouts
)
for chunk in stream:
    delta = chunk.choices[0].delta if chunk.choices else None
    if delta and delta.content:
        print(delta.content, end="")

1. Create a key in the OmniaKey dashboard and paste it in place of YOUR_OMNIAKEY_KEY.
2. Run it with streaming on: our parser task took almost three minutes.
3. To compare models, change only model to deepseek-flash or deepseek-v4-pro.

Conclusion

Kimi K3 is a strong, accurate open model: it passed every task we gave it. For everyday API work, though, DeepSeek V4.1 Flash gave the same answers 2 to 4 times faster and at a small fraction of the price, so start there. Reach for Kimi K3 when you need open weights at frontier scale or an agent that browses and calls tools all day.

Common questions

Is Kimi K3 open source?

Its full weights are open under the Kimi K3 License, an MIT-style license. Companies selling it as a hosted service with more than $20M in revenue over 12 months need a separate agreement with Moonshot, and very large products must show "Kimi K3" branding.

How much does Kimi K3 cost?

On Moonshot's international platform, $3.00 per million input tokens, $0.30 for cached input and $15.00 for output. The China platform charges ¥20, ¥2 and ¥100 for the same. There are no context-length tiers.

Can you turn off Kimi K3's thinking?

No. Moonshot's docs say K3 always thinks; you can only set the effort to low, high or max.

Is Kimi K3 better than DeepSeek V4 Pro for coding?

On our parser task both passed all 30 tests, but Kimi K3 took 176 seconds and V4 Pro 20.7 minutes, at about the same cost. Try a few of your own coding tasks on both before you switch.