DeepSeek vs Kimi (2026): Which Open-Weight Model Wins?

You are choosing between two Chinese AI labs’ open-weight model families for a real workload (an agent, a coding pipeline, a long-document RAG system) and the marketing pages are not helping. The deepseek vs kimi question has moved quickly in 2026. DeepSeek’s API now serves deepseek-flash (DeepSeek-V4.1-Flash-1017, released 17 October 2026) and deepseek-v4-pro (DeepSeek-V4-Pro-0813, released 13 August 2026), both with a 1M-token context and MIT-licensed weights. Moonshot AI’s Kimi API now leads with the K3 series: Kimi K3, which went live on the Kimi API in July 2026, and Kimi K3.1 Pro, the newest flagship on Moonshot’s pricing page, both with a 1,048,576-token context. This article gives you the verdict, the numbers, a worked cost example and a decision rule.

Verdict: which one to pick

For most cost-sensitive workloads, deepseek-flash is the cheaper default. At $1.20 per 1M output tokens at peak ($0.60 off-peak) it costs about an eighth of Kimi K3.1 Pro’s $10.00 output rate at peak, it accepts images, and you can turn thinking off or set its effort per request. Pick Kimi when you want Moonshot’s agent tooling (Kimi Code CLI, which Moonshot names as K3’s preferred agent framework), video input, or the models Moonshot positions for long-horizon coding and knowledge work, and you accept higher per-token prices.

Both families publish open weights, both expose OpenAI- and Anthropic-compatible APIs, and both now offer a context of about 1M tokens, so the old “DeepSeek for long context” argument no longer separates them. The split is mainly price, licence terms, modality and where your data is processed.

At a glance: deepseek vs kimi specs

From DeepSeek’s pricing page and Moonshot’s Kimi API pricing page and model cards, checked 26 October 2026. Prices are USD per 1M tokens.

Kimi’s K3-series prices are flat (no peak/off-peak split), and cache writes are billed separately by TTL: $2.00 (5-minute) or $4.00 (1-hour) per 1M tokens for K3.1 Pro, $3.00 or $6.00 for K3. Moonshot also lists a faster kimi-k3.1-pro-highspeed at $4.00 input / $20.00 output. The older K2-series models (K2.6 at $0.95 / $4.00, K2.7 Code) have a 262,144-token context; Moonshot says they will be unavailable to newly registered users from 31 October 2026 and plans to retire them by the end of December 2026.

Coding: where each model earns its keep

Both labs have leaned hard into agentic coding. The cleanest comparison is the one in DeepSeek’s own DeepSeek-V4-Pro-0813 model card (13 August 2026), whose Kimi K3 figures match the ones Moonshot publishes on the Kimi K3 model card:

On those rows Kimi K3 is ahead on DeepSWE and Toolathlon-Verified and slightly ahead on Terminal Bench 2.1 and HLE without tools, while DeepSeek-V4-Pro-0813 leads HLE with tools. DeepSeek’s later DeepSeek-V4.1-Flash-1017 card (17 October 2026) uses different benchmark versions and reports, among others, 77.0 for V4.1-Flash-1017 against 67.5 for K3 (listed as “K3”) on DeepSWE v1.1, and 68.2 against 58.0 on NL2Repo-Bench. Do not put the two cards’ numbers side by side: the versions and harnesses differ. We have not included K3.1 Pro scores: as of 26 October 2026, Moonshot’s K3.1 Pro quickstart shows its results only as charts, and its text still carries placeholder values.

Kimi’s distinct angle is agent orchestration. The Kimi K2.6 model card (April 2026) described an agent swarm scaling to 300 sub-agents executing 4,000 coordinated steps, and the K3 card describes long engineering sessions with minimal oversight; Moonshot recommends its Kimi Code CLI as K3’s agent framework. DeepSeek’s counter is price and its own integrations: its API docs list guides for tools such as Claude Code, OpenCode and GitHub Copilot, and FIM (Fill-In-the-Middle) completion is available as a Beta feature in non-thinking mode. For coding-specific trade-offs, see DeepSeek for coding.

Vision and multimodal coding

Both families now accept images. On DeepSeek’s API, deepseek-flash takes images in the content array as image_url blocks (base64 data URL or a public http(s) URL; JPEG, PNG, GIF and WebP), while deepseek-v4-pro does not support vision. Moonshot describes Kimi K3 as natively multimodal across text, images and video; its K3 quickstart accepts images as base64 or uploaded file IDs (not public image URLs) and video through its Files API. If your workflow needs video understanding, Kimi has the broader input support.

Reasoning: thinking modes compared

Both families let you set reasoning effort per request. The mechanics differ.

DeepSeek. Thinking mode is a request parameter, not a separate model ID, and it is on by default (effort high) on both deepseek-flash and deepseek-v4-pro. Set reasoning_effort to "low", "high" or "max", or pass extra_body={"thinking": {"type": "disabled"}} to opt out. The response returns reasoning_content alongside the final content; in thinking mode with tools, the reasoning_content of earlier turns must be passed back.

Kimi K3 and K3.1 Pro. Per Moonshot’s quickstarts, both always run with thinking enabled and take a top-level reasoning_effort of low, high or max (default max). Streaming returns separate reasoning_content and content deltas, and for multi-turn conversations and tool calls Moonshot asks you to send back the complete assistant message, not just content. The K3 quickstart also fixes temperature at 1.0 and top_p at 0.95. If you need a fast non-thinking path, DeepSeek’s opt-out is the simpler option.

Writing

Both labs build bilingual English and Chinese models, and both now accept prompts of about a million tokens, so a task like “rewrite this 8K-word document in a different register” fits either one with plenty of reference material and no chunking. Style is a matter of prompting and taste rather than published benchmarks, so try both on your own documents. On cost, long writing jobs are output-heavy, which is where deepseek-flash’s $1.20 peak / $0.60 off-peak output rate matters most. See DeepSeek for writing for prompt patterns.

Pricing: the headline reason most teams pick one

API pricing is where the gap opens. All figures are USD per 1M tokens from each company’s own pricing page, checked 26 October 2026.

DeepSeek’s peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays; all other hours are off-peak at half price. Kimi’s K3-series rates are the same at all hours, and Moonshot adds a cache-write charge each time a prefix is written to its cache. Moonshot says Kimi K3.1 Pro is unlocked after a minimum $1 top-up. For DeepSeek’s full rate card, see DeepSeek API pricing.

Worked example: 1M API calls per month

Take a SaaS app with a 2,000-token cached system prompt, a 200-token user message (uncached) and a 300-token response, run a million times in a month: 2,000M cached input tokens, 200M uncached input tokens and 300M output tokens.

                       deepseek-flash                 deepseek-v4-pro
                       PEAK        OFF-PEAK           PEAK        OFF-PEAK
Cached  2,000M tokens  × $0.006 =     $12.00  × $0.003 =      $6.00   × $0.044 =     $88.00  × $0.022 =     $44.00
Uncached  200M tokens  × $0.30  =     $60.00  × $0.15  =     $30.00   × $1.32  =    $264.00  × $0.66  =    $132.00
Output    300M tokens  × $1.20  =    $360.00  × $0.60  =    $180.00   × $3.96  =  $1,188.00  × $1.98  =    $594.00
Total                                $432.00              $216.00                $1,540.00              $770.00

The same workload on Kimi K3.1 Pro: 2,000M × $0.20 = $400.00; 200M × $2.00 = $400.00; 300M × $10.00 = $3,000.00; total $3,800.00, plus cache-write charges when the prefix is written. On Kimi K3 it is $600.00 + $600.00 + $4,500.00 = $5,700.00.

So for this token shape Kimi K3.1 Pro costs about 8.8× deepseek-flash run entirely at peak ($3,800 ÷ $432) and about 17.6× run entirely off-peak ($3,800 ÷ $216); against deepseek-v4-pro the ratios are about 2.5× (peak) and 4.9× (off-peak). Build your own scenarios with the DeepSeek pricing calculator.

Privacy and data residency

The two platforms differ here. DeepSeek’s Privacy Policy (last updated 10 February 2026) says it collects, processes and stores personal data in the People’s Republic of China. Moonshot’s Kimi Open Platform Privacy Policy says the platform is provided by Moonshot AI Pte. Ltd. in Singapore and that it stores the information it collects on servers in Singapore. Moonshot also offers Zero Data Retention to enterprise customers on request and says it does not use enterprise customer data to train models by default.

The strongest mitigation is the same for both: run the open weights yourself, so that prompts never leave your infrastructure, or limit what you send. Our DeepSeek privacy guide covers the DeepSeek side in detail.

Ecosystem and developer experience

DeepSeek’s API is OpenAI-compatible at https://api.deepseek.com; chat requests go to POST /chat/completions. The Anthropic SDK works against https://api.deepseek.com/anthropic, and the OpenAI Responses API format is also supported. The API is stateless: every request must resend the full conversation history. DeepSeek’s Privacy Policy describes a chat history that web and app users can manage, copy or delete in their settings; the API keeps no conversation state for you.

Minimal Python:

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ.get("DEEPSEEK_API_KEY"),
    base_url="https://api.deepseek.com",
)
resp = client.chat.completions.create(
    model="deepseek-flash",
    messages=[{"role": "user", "content": "Summarise this doc."}],
    reasoning_effort="high",
    extra_body={"thinking": {"type": "enabled"}},
    max_tokens=4096,
)
print(resp.choices[0].message.reasoning_content)  # reasoning_content
print(resp.choices[0].message.content)            # the final answer

Walk through it in the DeepSeek API getting started guide. DeepSeek limits concurrency per account to 2,500 requests on deepseek-flash and 500 on deepseek-v4-pro.

Kimi’s API follows the same OpenAI conventions at https://api.moonshot.ai/v1, and Moonshot also documents an Anthropic Messages-compatible API and a guide for using Kimi in Claude Code. Its rate limits (concurrency, RPM, TPM, TPD) depend on an account tier set by cumulative top-up. On top of the API, Moonshot offers Kimi Code CLI for terminal coding workflows.

Core parameters worth setting on DeepSeek: max_tokens, tool calling, JSON mode and streaming; temperature (DeepSeek recommends 0.0 for code, 1.3 for general chat, 1.5 for creative writing) only takes effect in non-thinking mode. JSON mode is designed to return valid JSON, not guaranteed: include the word “json” plus an example schema in the prompt and set max_tokens high enough to avoid truncation. See DeepSeek API JSON mode for the failure patterns.

When to pick which

The short version: DeepSeek on price and licence terms, Kimi on video input, Moonshot’s agent tooling and Singapore data storage.

Pick DeepSeek when

• Cost per token dominates your decision: deepseek-flash output is $1.20 peak / $0.60 off-peak against $10.00 for Kimi K3.1 Pro.
• You can shift batch work into DeepSeek’s off-peak hours for half-price tokens.
• You want a cheap model with image input (deepseek-flash) and a second model (deepseek-v4-pro) under the same API.
• You want MIT-licensed weights without revenue thresholds or user-interface display conditions, and you benefit from DeepSeek context caching on repeated prefixes.

Pick Kimi when

• You need video input as well as images.
• You want Moonshot’s agent stack (Kimi Code CLI) and the K3-series models it is built around.
• Your compliance team prefers a provider whose API data is stored in Singapore, or you need Zero Data Retention under an enterprise agreement.
• Your workload resembles the agentic benchmarks where Kimi K3 leads in DeepSeek’s own V4-Pro-0813 table (DeepSWE, Toolathlon-Verified).

Alternatives worth considering

If neither fits cleanly, the closest comparisons are DeepSeek vs Qwen, DeepSeek vs GLM (Z.ai’s GLM family) and DeepSeek vs MiniMax. For a wider field, see the DeepSeek comparisons hub.