.jpeg){kind=small}
If you want the lowest cost per token for analysis on your own marketing data, run DeepSeek V4 Pro. If you want the strongest data-handling terms and the fastest responses, run GLM-5.2. Kimi K3 scores highest on the independent Artificial Analysis index, but its API costs several times more and its weights carry a revenue-threshold license. All three publish weights that run on vLLM and SGLang.
This is the DeepSeek V4 vs Kimi K3 vs GLM-5.2 comparison written for the team that owns the marketing data, not for the benchmark leaderboard. It covers what each model costs per million tokens, what its license lets a company do, how much campaign history fits in one prompt, whether it can call your tools and return strict JSON, and what happens to your data when you use the vendor's API. Every outside claim was checked against the vendor's own pages or a named independent source on September 3, 2026; where the sources disagree, the disagreement is stated rather than resolved.
Key Takeaways
- DeepSeek V4 Pro is the cheapest of the three on the vendor API: $0.66 to $1.32 per million input tokens and $1.98 to $3.96 per million output tokens, depending on the hour and weekday, under an MIT license.
- Kimi K3 costs $3.00 per million input tokens and $15.00 per million output tokens on Moonshot's API, and its custom license requires a separate commercial agreement once the licensee's revenue, with its affiliates, exceeds 20 million US dollars over any consecutive 12 months.
- GLM-5.2 is the only one of the three whose privacy policy states that API content is not stored at all, and one third-party test measured it at roughly 168 tokens per second against about 62 for each of the other two.
- All three accept about 1 million input tokens; among the three, only Kimi K3 reads images and video.
- None of the vendors' headline benchmark numbers has an independent replication. The comparable third-party reading is Artificial Analysis's Intelligence Index: Kimi K3 60, GLM-5.2 53, DeepSeek V4 Pro 53.
What is an open-weight model, and why does it matter for marketing data?
An open-weight model is a large language model whose trained parameters are published for download, so a company can run it on its own hardware or through any inference provider instead of only through the vendor's API. For a marketing team the difference is where the data goes: with open weights, campaign and customer data can stay inside your own environment.
The three models in this comparison are the three largest open-weight releases of the summer: Moonshot AI unveiled Kimi K3 on July 16, 2026, according to Fortune's coverage (Moonshot's own announcement page was not found, so the weights release date is taken from the Hugging Face card rather than a dated post); Z.ai published GLM-5.2 in mid June 2026; and DeepSeek rolled out the GA release of DeepSeek-V4-Pro on its app, web and API on August 13, 2026, according to its own release notes. All three are mixture-of-experts models with total parameter counts in the trillions but far smaller active counts per token, which is why they can be served at all.
Marketing data adds three requirements that leaderboard comparisons skip. The model has to read long, structured inputs such as a month of daily spend by campaign across every channel. It has to return machine-readable output so a pipeline can trust it. And the terms of use have to survive a review by whoever owns customer data at your company.
How do Kimi K3, GLM-5.2 and DeepSeek V4 Pro compare at a glance?
DeepSeek V4 Pro wins on price and on the clarity of its MIT license, GLM-5.2 wins on speed and on data terms, and Kimi K3 wins on the independent intelligence score and on multimodal input. The table lists the facts as published by each vendor, with the source type noted where a number could only be confirmed independently.
| Attribute | Kimi K3 (Moonshot AI) | GLM-5.2 (Z.ai) | DeepSeek V4 Pro (DeepSeek) |
|---|---|---|---|
| Parameters (total / active) | 2.8T / 104B | 744B or 753B / 40B (the vendor's GitHub and Hugging Face pages differ) | 1.6T / 49B |
| Context window | 1,048,576 tokens | 1,000,000 input, 131,072 output | 1,000,000 tokens |
| Modalities | Text, image, video | Text only | Text only |
| Weights license | Custom Kimi K3 License (revenue and attribution thresholds) | MIT on the Hugging Face LICENSE file; Apache-2.0 in the GitHub README | MIT |
| Vendor API price, input / output per 1M tokens | $3.00 / $15.00 (cache hit $0.30) | $1.40 / $4.40 (independently measured; the official pricing page was unreachable) | $0.66 to $1.32 / $1.98 to $3.96 (off-peak / peak) |
| Inference stacks named by the vendor | vLLM, SGLang, TokenSpeed | vLLM, SGLang, Transformers, KTransformers, Unsloth | vLLM, SGLang, Transformers, Docker Model Runner |
| Tool calling and JSON output | Tool choice and dynamic tool loading; a JSON mode is not documented on the model card | Tool invocation, structured JSON output, MCP integration | OpenAI-compatible function calling (up to 128 functions, parallel, strict mode) and a JSON output mode |
| Data terms on the vendor API | Content may be used to improve the service; opt-out through enterprise agreements; servers in Singapore | Content is not stored; processed in real time in Singapore | Users may opt out of training use; retention "as long as necessary" |
| Artificial Analysis Intelligence Index | 60 | 53 | 53 |
How we evaluated
This comparison was built from each vendor's own documentation and from named independent sources, all fetched on September 3, 2026. Improvado does not sell any of these models. We make a marketing data platform that these models can read from, so our interest is in whether a model is safe, cheap and reliable enough to point at your data, not in which vendor wins.
Sources by claim type: parameters, context windows, modalities, licenses and inference stacks come from the Hugging Face model cards and GitHub repositories of moonshotai/Kimi-K3, zai-org/GLM-5.2 and deepseek-ai/DeepSeek-V4-Pro. Pricing comes from platform.kimi.ai and api-docs.deepseek.com; Z.ai's pricing page was unreachable, so GLM-5.2's price is the figure Artificial Analysis measured. Capability scores come from Artificial Analysis and from the CAISI assessment of GLM-5.2 published by the US National Institute of Standards and Technology. Data terms are quoted from each vendor's terms of service or privacy policy.
One honest limitation per model. Kimi K3's exact weights release date could not be confirmed on a Moonshot-owned page, and two of its most quoted benchmark claims appear only on aggregator sites. GLM-5.2's own vendor pages disagree with each other on license (MIT versus Apache-2.0), on parameter count (753B versus 744B) and on release date (June 13, 16 or 17); check the LICENSE file in the repository you actually download. DeepSeek V4 Pro's model card gives no GPU memory guidance for self-hosting, and a widely repeated "$0.435 / $0.87" price for it does not match any tier on DeepSeek's official pricing page. Not one of the three vendors' headline benchmark numbers has an independent replication of the exact figure.
Which model is cheapest to run on marketing data?
DeepSeek V4 Pro, by a wide margin on the vendor API. Its official price list charges $0.66 per million cache-miss input tokens and $1.98 per million output tokens off-peak, doubling to $1.32 and $3.96 during the peak window of 01:00 to 04:00 and 06:00 to 10:00 UTC. Cache-hit input is $0.022 to $0.044 per million. Kimi K3 charges $3.00 and $15.00 for the same units. GLM-5.2 sits between them at $1.40 and $4.40 as measured by Artificial Analysis.
The peak and off-peak split is the detail that matters for a reporting job. Most marketing analysis runs on a schedule you control, and DeepSeek's pricing page states that peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday, with all other hours billed off-peak. A nightly job that scores yesterday's campaigns at 23:00 UTC pays half of what an interactive session at 08:00 UTC on a weekday pays, and a weekend batch is off-peak all day.
At today's list prices, one dollar buys about 505,000 output tokens from DeepSeek V4 Pro off-peak (253,000 at peak), about 227,000 from GLM-5.2 and about 67,000 from Kimi K3; those are simple divisions of the per-million prices above. The ordering is not new: in a July 18, 2026 comparison of the three, MarkTechPost's Michal Sutter put DeepSeek V4 Pro at roughly 1.15 million output tokens per dollar and a blended per-task cost of $0.04 against $0.32 for GLM-5.2 and $0.94 for Kimi K3. His DeepSeek figure predates the peak and off-peak pricing that DeepSeek's release notes say took effect at 16:00 UTC on August 16, 2026, which is why it is about twice today's off-peak yield; the ranking and the size of the gap survive the change. Artificial Analysis independently rates both Kimi K3 and GLM-5.2 as expensive against the open-weight median input price of $0.50 per million tokens that its model pages showed beside each price on September 3, 2026 (a price median, distinct from the Intelligence Index median quoted later).
Speed changes the cost picture for interactive use. The MarkTechPost comparison measured GLM-5.2 at about 168 tokens per second against about 62 for each of the other two, and Artificial Analysis separately measured Kimi K3 at 38 tokens per second against the 64.5 median its model page showed on the same date. Throughput depends on the provider and the day, so the two sources need not agree on Kimi K3's exact number; both put GLM-5.2 well ahead. An analyst waiting on a chat answer feels the difference; a batch job does not.
Which license lets a marketing team use the weights?
DeepSeek V4 Pro is MIT-licensed, which permits commercial use, modification and redistribution with attribution and no revenue conditions. GLM-5.2 is also permissive, but Z.ai's own pages disagree on which permissive license applies: the LICENSE file in the Hugging Face repository reads "MIT License, Copyright (c) 2026 Zhipu AI", while the README of the official GLM-5 GitHub repository says Apache-2.0. Both permit commercial use; Apache-2.0 adds an explicit patent grant. Read the file in the repository you download and keep a copy with your deployment record.
Kimi K3 is different. It ships under a custom "Kimi K3 License" whose text requires a separate commercial agreement with Moonshot when "the aggregate revenue of the Licensee and its affiliates exceeds 20 million US dollars" over any consecutive 12 months, and requires the phrase "Kimi K3" to be prominently displayed by products with more than 100 million monthly active users or more than 20 million US dollars in monthly revenue. Internal use is exempt, as is access through Moonshot's official products or certified inference partners, and the weights are provided "as is" without warranty. For most companies above the revenue threshold that means the practical route to Kimi K3 is the API or a partner, not a self-hosted copy.
How much campaign history fits in one prompt?
Whether a quarter of daily, campaign-level performance data fits in one prompt depends on how many columns and channels the export carries; with about one million input tokens accepted by all three models it usually will, and the way to know is to test your own file. Kimi K3 lists 1,048,576 tokens on its model card, GLM-5.2 lists one million input tokens with output capped at 131,072, and DeepSeek V4 Pro lists one million. The difference between them is what the prompt can contain and how long the answer can be.
Kimi K3 is the only one of the three that reads images and video natively. That matters if the analysis includes creative: an agent that reviews the ad itself alongside its spend and conversion rows needs a multimodal model, and neither GLM-5.2 nor DeepSeek V4 Pro accepts anything but text. For pure tabular analysis the text-only models lose nothing.
GLM-5.2's 131,072-token output cap is generous for a report but worth knowing if you ask the model to rewrite a large dataset rather than summarize it. Long-context accuracy is a separate question from long-context capacity, and DeepSeek is the only vendor that publishes a long-context benchmark on its card, LongBench-V2 at 51.5, without independent replication. Test your own workload: a month of your real data, a question you already know the answer to, and a check of whether the model finds the right row.
Do they support tool calling and structured output for a data pipeline?
All three support tool calling, and two of the three document a strict JSON output mode. DeepSeek's documentation is the most specific for pipeline work: OpenAI-compatible function calling with up to 128 functions, parallel calls, a strict mode that "strictly adheres to the format requirements of the Function's JSON schema", and a separate JSON output mode that guarantees a valid JSON string. GLM-5.2's documentation lists tool invocation, structured output in JSON, streaming, context caching and MCP integration. Kimi K3's model card documents tool choice, dynamic tool loading and a preserved thinking mode, but it does not document a JSON mode, so plan on validating its output yourself.
For a marketing team the relevant pattern is the same for all three: the model does not hold the data; it generates a tool-call specification that the agent runtime interprets and executes against the system that does. A query tool that runs against a governed warehouse, a schema the model can read, and a strict output contract are what make the results reproducible. The model choice then becomes a question of cost, speed and terms rather than of who wrote the best demo.
What happens to your data on the vendor API?
The three vendors take three different positions, and the sentences are worth quoting. Moonshot's model-use terms state that "We may use Content to provide, maintain, develop, support, and improve the Services", with opt-out available through enterprise arrangements or separate written agreements; Kimi's privacy policy says information is stored on servers in Singapore. Z.ai's privacy policy states that the company does "not store any of the content the Customer or its End Users provide or generate while using our Services", that content is processed in real time and not saved, and that processing generally happens in Singapore. DeepSeek's privacy policy gives users "the right to opt-out of using your Personal Data for training our models", which implies training use by default, and retains personal data "for as long as necessary to provide our Services" without a fixed deletion period.
If customer-level marketing data is in scope, that reading favors GLM-5.2 among the hosted APIs, or self-hosting any of the three. Self-hosting is where the documentation runs out: as of September 3, 2026, none of the three model cards we reviewed states the GPU memory the full model needs, and the community figures that circulate are for smaller variants. Budget for a proper sizing exercise with your infrastructure team before promising anyone a self-hosted deployment, and treat the vendor API as the way to evaluate the model, not as the final architecture.
How capable are they, and who checked?
On the one third-party index that scores all three the same way, Kimi K3 leads with 60 on Artificial Analysis's Intelligence Index, and GLM-5.2 and DeepSeek V4 Pro tie at 53, all well above the open-weight median Intelligence Index score of 29 that Artificial Analysis showed beside each model's score on September 3, 2026. The US government's CAISI assessment, published by NIST on July 17, 2026, states that GLM-5.2's overall capabilities "are similar to that of GPT-5.2, released in December 2025", in NIST's own words, and also found that its safeguards block fewer sensitive requests than the US reference models it was compared with.
The vendors' own numbers are higher and narrower. Moonshot's card lists GPQA Diamond 93.5 and BrowseComp 91.2 for Kimi K3; Z.ai's card lists SWE-bench Pro 62.1 and GPQA Diamond 91.2 for GLM-5.2, with a Terminal-Bench figure that differs by 1.7 points between its two official pages; DeepSeek's card lists MMLU-Pro 87.5 and LiveCodeBench 93.5 for V4 Pro. None of these exact figures has been reproduced by an independent party in the sources we could find, which is the same caveat independent reviewers such as Simon Willison have attached to this generation of releases. Read the vendor numbers as the vendor's claim and the Artificial Analysis index as the comparable measurement.
How Improvado fits
The model is the easy part of this decision. The hard part is the data it reads: spend from a dozen ad platforms, conversions from the CRM, revenue from the warehouse, all with different names for the same campaign. Improvado collects that marketing data, harmonizes it into one governed model in your warehouse, and exposes it to AI agents through an MCP server, so the model you choose queries the same numbers your dashboards show instead of a spreadsheet export. Because the data layer stays yours, switching from one model to another is a configuration change, not a migration.
If you are deciding which of these models to put in front of your marketing data, talk to an expert about the data side first. The comparison above tells you what each model costs and permits; the pipeline decides whether its answers are right.
Frequently Asked Questions
Which is cheaper for marketing analysis, DeepSeek V4 Pro or Kimi K3?
DeepSeek V4 Pro. On the vendors' own price lists DeepSeek charges $0.66 to $1.32 per million input tokens and $1.98 to $3.96 per million output tokens depending on the hour, while Kimi K3 charges $3.00 and $15.00. A batch job scheduled outside DeepSeek's peak window pays the lower tier.
Can a company with more than 20 million dollars in revenue self-host Kimi K3?
Only with a separate commercial agreement. The Kimi K3 License requires one when a licensee's aggregate revenue exceeds 20 million US dollars over any consecutive 12 months; internal use and access through Moonshot's official products or certified inference partners are exempt from that clause.
Is GLM-5.2 MIT or Apache-2.0 licensed?
Z.ai's own pages disagree: the LICENSE file in the Hugging Face repository says MIT, and the README of the official GitHub repository says Apache-2.0. Both permit commercial use. Keep a copy of the license file from the repository you download with your deployment record.
Which of the three models does not store API data?
GLM-5.2. Z.ai's privacy policy states that customer content is processed in real time and not saved on its servers. Moonshot's terms allow content to be used to improve the service unless an enterprise agreement says otherwise, and DeepSeek's policy offers an opt-out from training use.
Do all three models accept a million tokens of context?
Yes. Kimi K3 lists 1,048,576 tokens, GLM-5.2 lists one million input tokens with a 131,072-token output cap, and DeepSeek V4 Pro lists one million. A quarter of daily campaign-level performance rows fits in one prompt on any of them.
Can these models return strict JSON for a data pipeline?
DeepSeek V4 Pro and GLM-5.2 document it: DeepSeek offers a JSON output mode and strict-schema function calling, and GLM-5.2 lists structured JSON output and MCP integration. Kimi K3 documents tool calling but not a JSON mode, so validate its output in your pipeline.
