Kimi K3 vs DeepSeek V4: two 1M-context models with different bills

Kimi K3 figures come from the official quickstart at platform.kimi.ai and the official chat pricing page. DeepSeek figures come from the official pricing and model pages at api-docs.deepseek.com. Prices change; confirm them before you commit a budget.

Both vendors publish a one-million-token context window, so a comparison built only on context length would end in a draw. The difference appears at the other end of the request. Kimi K3 documents a maximum completion of 1,048,576 tokens, with a default of 131,072. DeepSeek V4 Pro publishes a 384K maximum output. If the job is a long generated file, a multi-file patch, or an extended research memo, that ceiling matters more than the input window.

A large context window is still not a context strategy. Duplicated repository files, full logs, and irrelevant attachments raise latency and cost even when the window accepts them. Both providers bill input tokens, so selecting the evidence is part of the job on either model. Neither vendor's context number removes that work.

Kimi K3 officially documents text, image, and video input with native visual understanding. DeepSeek's current pricing and model pages describe text models and do not publish an image-input contract. That makes the decision simple for a workflow built on screenshots, charts, interface debugging, or video frames: verify that the provider accepts the input at all before comparing quality or price. This page does not claim that DeepSeek can never process an image; it reports what the official documentation reviewed on 21 October 2026 actually says.

On the K3 Kimi hosted chat, image input has its own price: a text response costs 100 credits and a response with one image costs 150. One image up to 4 MB is accepted per message. For a multi-image or video workflow, the API is the appropriate path, not the hosted chat.

Kimi K3 uses flat pay-as-you-go pricing with no tiering by context length. The official chat pricing page lists $3.00 per million input tokens on a cache miss, $0.30 per million on a cache hit, and $15.00 per million output tokens. Cache writes are billed separately: $3.00 per million tokens for a five-minute TTL and $6.00 for one hour. Thinking is always on, and reasoning tokens are billed as output.

DeepSeek V4 Pro lists $1.32 per million input tokens on a cache miss, $0.044 on a cache hit, and $3.96 per million output tokens at peak. Off-peak rates are exactly half: $0.66 input, $0.022 cache hit, and $1.98 output. Peak hours are 01:00-04:00 and 06:00-10:00 UTC, Monday through Friday, excluding Chinese public holidays. Every other hour, including weekends, bills at the lower rate. DeepSeek Flash is cheaper again at $0.30 input and $1.20 output at peak, with the same half-price off-peak rule.

Run the arithmetic on a representative request instead of comparing headline rates. For 100,000 input tokens on a cache miss and 10,000 output tokens, Kimi K3 lists $0.30 plus $0.15, or $0.45. DeepSeek V4 Pro lists about $0.17 at peak and about $0.09 off-peak for the same request. DeepSeek Flash lists about $0.04 at peak. None of these figures includes retries, reasoning-token overhead, or the cost of the human time spent checking the answer.

The Kimi K3 quickstart states that the full model weights have been released and describes the model as the first open-source model in the 3-trillion-parameter class. That is useful for teams that need to inspect or self-host a model, but it is not a cheap deployment: 2.8 trillion parameters implies substantial hardware, and the documentation fixes sampling settings and keeps thinking on. Open weights are a control and audit advantage, not a discount.

The DeepSeek model page reviewed for this comparison does not state weight availability. Do not assume either direction from a third-party leaderboard; check the vendor's own repository and licence before planning a self-hosted deployment. For most teams, the practical choice on both sides is the hosted API, and the decision comes back to modality, output length, and price.

Choose Kimi K3 when the task needs image or video understanding, a very long generated answer, released weights, or a single model that covers text and visual reasoning. It is also the right choice when you want to chat with K3 without managing an API account: K3 Kimi provides hosted access with fixed per-response credits, so the cost is visible before you send.

Choose DeepSeek V4 when the workload is text, the volume is high, and cost per token dominates. Off-peak scheduling can halve the bill, and the 384K output ceiling is generous for most code and document tasks. DeepSeek Flash is the stronger fit for high-volume, latency-tolerant batch work where the extra capability of a larger model is not needed.

Neither model is the right answer for a task that needs a deterministic calculation or a citation you have not checked. Both are probabilistic models; a lower price does not make a wrong answer cheaper to fix. For a plain chat with K3, start on the K3 Kimi homepage. For plan and credit pricing, see the pricing page. For the earlier generation comparison, read Kimi K3 vs Kimi K2.

This page does not publish a head-to-head benchmark of Kimi K3 against DeepSeek V4. We have not run a controlled evaluation on the same tasks, and we will not present a vendor's marketing chart as one. Published benchmark scores, including the values on the Kimi K3 benchmarks page, are selection signals. They do not guarantee that a specific coding, research, or extraction task will work.

The honest evaluation is a small, representative test on your own inputs: the same ten prompts, the same documents, the same review criteria, and a recorded cost per accepted answer. That number will be more useful than any list price, including the ones quoted here.