Qwen4-Max joins a choice of Chinese-developed models that already includes Kimi K3.1 Pro, GLM-5.3 and DeepSeek V4 Pro. The useful comparison is more specific than a leaderboard: which exact model can your team deploy, through which provider, under which terms, and with what behaviour in the application?
This guide uses primary documentation, actual licence files and evaluator pages checked on November 17, 2026. It compares Qwen’s production model, Kimi K3.1 Pro, GLM-5.3 and DeepSeek V4 Pro 0813. A similarly named older DeepSeek router listing is treated separately. No new inference benchmark or self-hosted deployment was performed; calculated examples are labelled as such.
- 01Specify model and provider togetherA family name can hide different revisions or service modes. Pair each quote and result with the exact route that produced it.
- 02Read the actual weight licenceKimi and GLM have custom commercial conditions, DeepSeek 0813 publishes an MIT licence, and Qwen4-Max has no public checkpoint in the reviewed release.
- 03Treat reasoning as part of the contractKimi and GLM always reason; other candidates permit non-thinking modes. Effort labels and continuation requirements are not universal.
- 04Evaluate the complete workflowMedia handling, tool state, cache charges and failed attempts determine the operating result more reliably than one input-token price.
01 — Exact identityDo not combine one revision’s price with another’s result
The current DeepSeek rate card maps deepseek-v4-pro to DeepSeek-V4-Pro-0813. The OpenRouter page with a similar slug identifies its model as DeepSeek V4 Pro 0423. At retrieval, that older listing included a Relace provider row at $0.90 input and $1.50 output per million tokens. It is not a quote for the direct 0813 service used in this comparison.
Kimi presents a different purchasing distinction. Its router listing showed a Relace offer around $0.2403 input and $8.4023 output per million tokens, while the direct standard API remained $2/$10. A provider-specific offer can be useful, but it is not evidence that Moonshot changed its direct rate. Record the selected provider and the offer’s current conditions before using it in a budget.
Create a short identity record for every candidate: model revision, provider, endpoint, service mode and retrieval date. Keep that record with the evaluation output. The DeepSeek 0813 release analysis provides historical context, but a production decision should use the current route. If an aggregator changes its chosen provider, the resulting service may need another compatibility and cost check even when the visible model name stays the same.
The benchmark, direct price and router quote must refer to the same revision and route before they can support a combined claim. DeepSeek 0423 and 0813 are separate comparisons.
02 — Deployment termsSeparate hosted access from permission to run the weights
Qwen4-Max has no public weights in the reviewed release. Kimi, GLM and DeepSeek have downloadable artifacts, but their licences differ. The Kimi K3.1 Pro licence requires a separate agreement when a licensee or affiliate operates a Model as a Service business and aggregate revenue exceeds $20 million over any consecutive 12 months. It also contains a separate branding condition for a commercial product above 100 million monthly active users or $20 million monthly revenue.
Kimi’s sections 2 and 3 have defined exceptions for internal use and access through official products or certified inference partners. Internal use excludes making the software, outputs or underlying capabilities available to third parties. The GLM-5.3 licence instead requires a security review for a qualifying Model as a Service operator whose licensee-and-affiliate aggregate revenue exceeds $10 billion over any consecutive 12 months. Both define that service category around meaningful third-party control, with exclusions for embedded features and mere request relaying.
The DeepSeek 0813 file is MIT, including its notice-preservation condition and warranty disclaimer. These are selected distinctions, not substitutes for the full terms of the artifact and deployment route. Our GLM licence guide explains why “downloadable” and “MIT” should not be used interchangeably. Resolve the applicable deployment model before investing in infrastructure or promising a customer a particular hosting arrangement.
03 — Direct ratesCompare standard calls before adding service-specific savings
The baseline below uses QwenCloud Singapore pricing, Kimi’s standard API, Z.ai’s rate card and DeepSeek’s direct 0813 rates. Kimi Highspeed is a separate $4/$20 input/output service; it is not the standard route. DeepSeek uses peak and off-peak pricing, so a single undated low rate is incomplete.
For an illustrative collection of separate valid requests totalling one million uncached input tokens and 100,000 billed output tokens, the token charges are $1.95 for Qwen, $3 for Kimi standard, $1.84 for GLM and $1.716 for DeepSeek peak or $0.858 off-peak. This holds token quantities constant to show rate differences. It does not claim the models use equal reasoning or achieve equal quality, and it excludes caches, tools, taxes and repeated attempts.
DeepSeek’s peak windows are 01:00–04:00 and 06:00–10:00 UTC on weekdays, excluding Chinese public holidays; other times are off-peak. If work can wait, scheduling may change its token bill, but deadlines and queue behaviour still matter. Use the arrival pattern of the real workload rather than assuming every request qualifies for the lowest rate. Subscription allowances and coding-plan points belong in a separate analysis from these pay-as-you-go charges.
| Direct route | Input / output per 1M | Cache read per 1M |
|---|---|---|
| Qwen4-Max, Singapore | $1.50 / $4.50 | $0.075 |
| Kimi K3.1 Pro, standard | $2.00 / $10.00 | $0.20 |
| GLM-5.3 | $1.40 / $4.40 | $0.26 |
| DeepSeek 0813, peak | $1.32 / $3.96 | $0.044 |
| DeepSeek 0813, off-peak | $0.66 / $1.98 | $0.022 |
04 — Media and limitsLet the required evidence narrow the candidate list
The Qwen GA specification supports text, images and video. The Kimi K3.1 Pro guide documents images and video as well as text, while GLM-5.3 and direct DeepSeek 0813 are text-only. If the job depends on visual evidence, a text-only candidate needs an additional extraction step whose errors and costs belong in the evaluation.
The large context windows also have different completion rules. Qwen GA allows 131,072 completion tokens without thinking or 262,144 with it. GLM lists 128K maximum output and DeepSeek 384K. Kimi defaults to 131,072 completion tokens and permits configuration up to 1,048,576, subject to the context available. None of these output figures should be added to a full input window as an independent allowance.
The upload path matters as much as the modality label. Kimi’s documented vision route uses base64 or file references rather than public image URLs, and its video example uses the Files API. A client built around URL-only image inputs therefore needs adaptation before a fair trial. Our Kimi K3.1 Pro release guide covers the broader release, while the production integration should be checked against the current provider contract.
05 — Reasoning contractTest continuation and effort controls as application behaviour
Kimi and GLM always operate with reasoning enabled and offer low, high and max effort, defaulting to max. GLM explicitly rejects disabling reasoning; its migration guidance suggests low effort when moving from a formerly non-thinking configuration. Qwen allows thinking and non-thinking modes, while DeepSeek defaults to thinking at high effort. A shared client setting called “high” does not establish equal compute or equivalent behaviour across these providers.
The DeepSeek thinking guide maps medium and xhigh to high, and max or ultra to max. When requests carry tools, it requires previous reasoning_content to be returned in later requests, including turns that did not invoke a tool; missing it can produce a 400 error. Kimi requires the complete assistant message to be retained across turns. Treat provider state as protocol data rather than reducing every message to the text displayed in the interface.
Exercise the workflow past its first answer. A model that responds correctly to a single prompt may fail when a tool returns an error, a user revises the request or the client resumes a session. The GLM thinking migration article shows why this can be a breaking change. Keep permissions and action validation in the application, and preserve only the state required for the chosen protocol without exposing internal reasoning as a product feature.
06 — Evaluator snapshotUse one index version and preserve the disclosed settings
Artificial Analysis pages checked November 17 use Intelligence Index v5.0.1 for Qwen4 Max, Kimi K3.1 Pro, GLM-5.3 Max and DeepSeek V4 Pro 0813 Max. The retrieved Qwen and Kimi summaries do not disclose an effort suffix. Keep that uncertainty rather than inferring a tested setting from the API default.
Kimi has the highest index score among these four in this snapshot. That observation does not identify the best provider for every workflow, and it should not be enlarged into an overall market ranking. The evaluator’s cost is a weighted average per index task and its time is an index-task measure. DeepSeek’s displayed rate assumptions use peak pricing, so its reported cost is not an off-peak forecast.
The table helps explain why direct rates alone are insufficient: GLM’s lower headline rates than Kimi do not produce a lower evaluator task cost in this snapshot. Task mix, reasoning and behaviour matter. Use the result to choose questions for a trial, such as whether a candidate completes a difficult job with fewer repairs. Avoid multiplying these benchmark costs by your business task count without measuring the workload that generates those tasks.
| AA configuration | Index v5.0.1 | USD / index task | Minutes / index task |
|---|---|---|---|
| Qwen4 Max, effort undisclosed | 43 | $3.27 | 32.2 |
| Kimi K3.1 Pro, effort undisclosed | 46 | $3.68 | 38.5 |
| GLM-5.3 Max | 32 | $5.01 | 22.2 |
| DeepSeek 0813 Max | 23 | $1.77 | 17.3 |
07 — Operating ledgerAdd cache and tool behaviour before predicting savings
Repeated-input prices are only one part of the ledger. Kimi’s standard route lists cache writes at $2 per million tokens for five minutes or $4 for one hour; Qwen lists $1.875 and $3 for those lifetimes. GLM’s cache storage is currently labelled limited-time free. Record that qualification rather than treating it as a permanent zero-cost feature. Use actual hit and write records to determine whether repeated context is helping.
Discounted modes have their own contracts. Kimi Batch is billed at 60% of standard rates, which means a 40% reduction, not the 50% reduction common in some other services. Its current quickstart also says the web-search tool is being updated and is not recommended for near-term production workflows. If search is necessary, resolve that dependency before interpreting a successful model-only example as a complete application trial.
Add the costs outside inference: preprocessing, tool execution, queues, storage and human repair. For self-hosted candidates, include the infrastructure and operating effort needed to serve the actual artifact; active parameter count is not a hardware bill. Our AI cost-optimisation playbook provides a wider accounting structure. Keep the direct and self-hosted estimates separate so a theoretical token saving does not conceal a new operational burden.
08 — Decision pathChoose a candidate for a specific reason, then test that reason
For a hosted visual workflow, Qwen and Kimi are candidates because their documented inputs include images and video. For a text-only workflow, all four can enter the evaluation, subject to access and integration constraints. If downloadable weights are mandatory, Qwen4-Max does not satisfy that requirement in the reviewed release. The remaining choices then need artifact-specific licence and infrastructure assessment before quality comparisons become decisive.
Make the proposed advantage explicit. It might be better acceptance on difficult coding tasks, a lower direct bill during eligible hours, or the ability to operate a licensed checkpoint within a required environment. Select examples that can falsify that advantage as well as demonstrate it. Measure total spend across successful and failed attempts, accepted results, elapsed time and reviewer effort. Keep the model revision and route attached to every result.
Our AI transformation services help turn those findings into a deployable choice. Prefer a small, reversible adoption step with an owner and a clear success condition. Expand only when the evidence supports the intended benefit. This comparison is most useful as a way to avoid mismatched assumptions and select a fair trial, rather than as a permanent declaration that one model wins every category.
Evaluate Qwen or Kimi
Check the upload path and visual evidence before comparing response quality.
Compare the full candidate set
Use task acceptance, total spend and elapsed time under explicit reasoning settings.
Assess the actual checkpoint
Separate downloadable artifacts from hosted services and account for the infrastructure.
Keep the identity, contract and outcome in one decision
A credible comparison names the exact revision, provider and commercial route before quoting a price or score. Qwen, Kimi, GLM and DeepSeek each deserve evaluation where their capabilities and terms fit the job. Preserve those distinctions, test the full workflow and choose from accepted results rather than combining the cheapest quote with the strongest unrelated benchmark.