QVeris
MODEL GATEWAY · NEW MODEL

Qwen4-Max: pricing, real-world cases and model comparison

Qwen4-Max is now on QVeris. Compare gateway pricing, explore official use cases, and connect a million-token model to your workflow.

Compare model capabilities

Updated November 17, 2026 · Created by the QVeris team

Qwen4-Max vs. leading models

Use the short inline trials below for a quick comparison. The five complete professional tasks are available in the Agent prompt and expandable task list.

Investment memo and auditable financial model

Act as a research analyst preparing an auditable investment-committee draft for fictional industrial-software company Atlas. Use only the supplied data; all amounts are CNY million. [F1 audited FY2025] revenue 480, recurring revenue 360, gross margin 72%, operating expenses 260, depreciation 12, capex 18, net working capital 50, net debt 60, diluted shares 100 million. [F2 FY2026 signed forecast] recurring revenue growth 18%, non-recurring revenue growth 5%, gross margin 73%, operating expenses 290, depreciation 14, capex 22, closing working capital 58, tax rate 25%; assume full use of tax benefits. [F3 management presentation] revenue growth 25%, without a supporting bridge. Define EBIT=revenue*gross_margin-opex; FCFF=EBIT*(1-tax)+depreciation-capex-change_in_working_capital. Deliver: (1) a source-linked FY2025-to-FY2026 revenue bridge, EBIT and FCFF; (2) a 3x3 sensitivity table using recurring growth 10%/18%/26% and gross margin 70%/73%/76%, holding all other FY2026 inputs fixed; (3) a spreadsheet specification with Inputs, Forecast, Sensitivity and Checks tabs, including cell addresses and actual Excel formulas; (4) a six-slide committee narrative with thesis, contradictory evidence, downside triggers and diligence requests. Distinguish audited facts, assumptions and calculations. Reconcile F2 with F3. Do not invent market multiples, prices or an executed workbook. Show formulas and units; round display values to two decimals.

Acquisition diligence and evidence reconciliation

Prepare an evidence-based diligence memo for a fictional software acquisition. Documents: [D1 executed customer contract, Jan 5] annual fee CNY 2.4m, 30-day termination for convenience, change of control requires written customer consent, liability capped at fees paid in the preceding 12 months; [D2 unsigned amendment, Mar 2] removes termination and consent requirements; [D3 revenue ledger, Jun 30] recognizes CNY 1.2m from this customer, representing 35% of YTD revenue; [D4 data-processing annex] storage permitted only in region A, subprocessors require prior written approval; [D5 deployment inventory] backups in region B, vendor Z processes support logs; [D6 management email] customer verbally agreed and vendor approval is being arranged; no written records attached. Produce (1) document register and authority/conflict matrix, (2) clause-to-evidence risk register with severity, source IDs, missing facts, owner and closing condition, (3) customer-concentration and revenue-at-risk scenarios without treating scenario exposure as booked loss, and (4) a 200-word deal-team memo distinguishing confirmed facts, unresolved issues and actions before closing. An unsigned amendment cannot be treated as effective. Verbal assertions are not written consent. Use only the contractual terms provided; do not assume jurisdiction, decide enforceability or fabricate legal conclusions. Cite each material assertion.

Repair a production webhook ingestion service

You own a Python 3.11/SQLite webhook ingester. Existing pseudocode: if not SELECT event_id from events: UPDATE orders SET status=payload.status; INSERT INTO events(event_id); commit(). Incident: concurrent deliveries duplicate side effects; a retry after a crash can leave the event marked processed without an order update; a delayed paid event overwrites refunded. Event contract: event_id string, order_id string, version nonnegative integer excluding bool, status in pending/paid/refunded. Legal transitions: pending->paid->refunded; repeated same status allowed. Greater version wins only if transition is valid; older/equal versions are recorded as stale without changing the order. Reusing an event_id with different canonical payload must be rejected. Orders already exist; unknown order means rollback. Deliver a minimal runnable standard-library implementation ingest(conn,payload), SQL schema, and unittest tests for duplicate delivery, payload mismatch, out-of-order events, illegal regression, unknown order and atomic rollback under injected failure. Use a transaction and uniqueness constraints; no check-then-write race. Return explicit applied/duplicate/stale outcomes. Explain SQLite concurrency/retry handling and why external side effects require an outbox rather than a claim of exactly-once delivery. Provide complete code; do not claim you ran tests or accessed a real repository.

Parametric 3D product asset and reproducible delivery

Create a reproducible Blender 4.x Python script for a fictional industrial centrifugal-pump product visualization. It must run in a blank scene using bpy, math and standard-library modules only, without downloaded assets. Parameters in meters: base_length=.8, base_width=.4, motor_radius=.12, motor_length=.28, housing_radius=.18, pipe_radius=.04. Build separately named base, motor, shaft, housing, inlet, outlet and mounting bolts; model pipes as hollow geometry rather than solid cylinders. Use distinct metal/rubber materials, bevels, a neutral studio ground, three lights and an orthographic camera framing the entire product. Make an exploded-view parameter with 0=assembled and 1=exploded; changing it must preserve each component's identity. Rerunning the script may replace only objects in its own named collection. Deliver complete runnable script, parameter table, output-saving instructions and five programmatic geometry/scene checks. Use explicit output paths supplied via script parameters; do not modify unrelated files or claim a rendered artifact. State that this is a visual approximation, not a manufacturing-ready CAD design. Keep topology and transform choices explainable.

Scientific screening and experimental validation design

Design the next validation round for fictional data-center cooling fluids. Data at 25°C: baseline B thermal conductivity .14 W/(m K), dynamic viscosity 2.0 mPa s, flash point 150°C, cost 20 CNY/kg; candidate A .17, 3.2, 165, 25; candidate C .16, 1.8, 140, 22. Each is a single measurement with no uncertainty estimate. Project requirements: flash point >=160°C, viscosity <=3.0 mPa s at 25°C, material compatibility and aging data required before any pilot; thermal conductivity is to be maximized only after these gates. A simulation predicts A reduces rack temperature by 8°C but uses fixed pump power, whereas the baseline test used fixed flow. Deliver (1) a requirement/evidence matrix and candidate disposition, (2) analysis of the simulation-versus-test confound, (3) a preregistered experiment plan with controls, randomized run order, repeat structure, sensor calibration, primary endpoint and exclusion rules, (4) a staged decision table for repeat laboratory testing, compatibility/aging review and a controlled engineering pilot, and (5) machine-readable JSON listing missing data and stop conditions. Separate measured values, simulation predictions and hypotheses. Do not invent error bars, sample-size power, safety certifications or a proven best candidate. Explain what additional information is required for a power calculation.

Model capabilities, pricing and per-call costs

Qwen4-Max is now listed in the QVeris Model Gateway. Use the same inputs and acceptance criteria to compare it with DeepSeek V4.1 Flash 1017, Kimi K3.1 Pro, Gemini 4 Carbon, GPT-6.1 Sol and Claude Fable 5.2. The evaluation focus below is a suggested test plan, not a performance ranking.

All six QVeris routes advertise reasoning, tool calling, JSON output, streaming and vision. These are supported features, not evidence of equal answer quality. The table shows route-specific input formats and token limits; maximum output is a configured ceiling and shares the context budget with inputs.

Model / evaluation focusInput → outputReasoning / tools / JSON / streamingContext / max output tokens
Qwen4-Max
Documents, visual reasoning and staged delivery
Image, text → textSupported / supported / supported / supported1,000,000 / 262,144
Gemini 4 Carbon
Cross-provider visual and long-context reference
Image, text → textSupported / supported / supported / supported1,048,576 / 943,718
GPT-6.1 Sol
Code correctness and structured task execution
Image, text → textSupported / supported / supported / supported1,050,000 / 128,000
Claude Fable 5.2
Complex reasoning and agent workflow review
File, image, text → textSupported / supported / supported / supported1,000,000 / 128,000
DeepSeek V4.1 Flash 1017
End-to-end project development and tool-driven iteration
Image, text → textSupported / supported / supported / supported1,048,576 / 393,216
Kimi K3.1 Pro
Coding, 3D creation and long-running research agents
Image, text → textSupported / supported / supported / supported1,048,576 / 1,048,576

How much does one call cost?

There is no flat per-call price. For a comparable example, assume each model receives 10,000 uncached input tokens and produces 2,000 billed output tokens, including any billed reasoning. Unit prices below are Credits per million tokens; the last column is the estimated Credits for that single call.

ModelInput / output priceCached input priceOne call / Credits
Qwen4-Max741.825 / 2,22637.0912511.87025
Gemini 4 Carbon1,050 / 5,25052.521
GPT-6.1 Sol1,050 / 5,25052.521
Claude Fable 5.24,200 / 21,00010584
DeepSeek V4.1 Flash 1017Request-dependentSee billing rulesCannot estimate from a fixed rate
Kimi K3.1 Pro1,015.9275 / 5,314.1025101.587520.78748

Calculation: 10,000 ÷ 1,000,000 × input price + 2,000 ÷ 1,000,000 × output price. Qwen4-Max costs 11.87025 Credits in this example. Sending the same test to Qwen4-Max and Kimi K3.1 Pro creates two model calls, totaling 32.65773 Credits under these assumptions.

Estimates exclude cache writes, additional tools and retries. Actual token counts vary by model. GPT-6.1 Sol uses higher rates from 272,000 input tokens; the example uses its first tier. DeepSeek does not publish a fixed route rate here. Credits are not currency amounts; check your purchase package and actual usage ledger. Each model name links to the corresponding QVeris pricing source.

DeepSeek and Kimi: current domestic alternatives

DeepSeek’s 1017 release focuses on complete-project development with feedback-driven iteration. Compare code repair, runnable project delivery and recovery after tool failures. Native Search & Fetch services should be validated separately from the gateway model route. DeepSeek release notes

Kimi describes K3.1 Pro as focused on coding, 3D creation, financial research and long-running agents. Compare cross-document evidence, executable scripts and staged deliverables; evaluate observed task results rather than transferring vendor benchmark rankings to your workload. Kimi model documentation

QVeris model parameters

ParameterCurrent QVeris route
Model IDqwen4-max
ProviderAlibaba
Context / maximum output1,000,000 / 262,144 tokens
Input → outputImage, text → text
CapabilitiesChat, JSON, reasoning, streaming, tools, vision
API operationschat.completions · messages · responses

This route currently lists image and text input. Native Qwen video support does not establish video support on the gateway. QVeris model details

QVeris Model Gateway: integration and usage

Use one QVeris API key and an OpenAI-compatible endpoint to evaluate multiple models. Create a key, keep it on your server, then check the current model list before integration. Gateway documentation

StepAction
01Choose qwen4-max from GET /v1/models.
02Configure https://aigateway.qveris.ai/v1 as base_url.
03Start with a small representative request, then expand task coverage.
04Retain the call ID and inspect qveris_billing and usage history.
Python · OpenAI SDK
import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["QVERIS_API_KEY"],
    base_url="https://aigateway.qveris.ai/v1",
)
response = client.chat.completions.create(
    model="qwen4-max",
    messages=[{"role": "user", "content": "Create a staged project plan."}],
    max_completion_tokens=4096,
)
print(response.choices[0].message.content)

Chat Completions, Responses and Anthropic-compatible Messages are documented gateway operations; model support still matters. Tool definitions require an application to execute tools and return results. Native hosted tools and optional provider parameters need separate route validation.

Multi-model API pricing and request costs

Snapshot checked November 17, 2026. All gateway rates below are Credits per million tokens, using the first input tier. The example uses 10,000 uncached input tokens and 2,000 billed output tokens, without extra tool charges.

ModelInputOutputExample / Credits
Qwen4-Max741.8252,22611.87025
Gemini 4 Carbon1,0505,25021
GPT-6.1 Sol1,0505,25021
Claude Fable 5.24,20021,00084
DeepSeek V4.1 Flash 1017Request-dependentRequest-dependentSee billing rules
Kimi K3.1 Pro1,015.92755,314.102520.78748

Qwen4-Max cached input is 37.09125 Credits / 1M tokens. With 8,000 of the 10,000 input tokens cached, the example becomes 6.23238 Credits. GPT-6.1 Sol lists tiered pricing; check its detail page for the applicable tier. Credits are not USD, and their purchase cost depends on the package. Credits pricing

Kimi K3.1 Pro uses the detail-page rates of 1,015.9275 input and 5,314.1025 output Credits / 1M tokens; cached input is 101.5875. Its uncached example costs 20.78748 Credits. DeepSeek V4.1 Flash 1017 publishes request-dependent gateway pricing, so no fixed rate or comparable estimate is invented. Kimi gateway pricing · DeepSeek billing details

Qwen native API pricing

Region / currencyInput / output per 1MCache readCache write 5m / 1h
China platform · CNY¥10 / ¥30¥0.5¥12.5 / ¥20
Singapore · USD$1.50 / $4.50$0.075$1.875 / $3.00

These native API rates are separate from QVeris billing and Token Plan subscriptions. The same uncached example is ¥0.16 on the China platform or $0.024 in Singapore. Output includes billed thinking tokens. No input-length tier is listed for Qwen4-Max within its supported window. China pricing · Regional USD pricing

Qwen4-Max strengths and use cases

For a useful pilot, choose work that can be checked: a sourced document summary, a repairable code change or a staged plan with clear acceptance criteria. Assess completeness, traceability and total task cost together.

Native multimodality and tool support

Qwen documents image and video understanding alongside reasoning, function calling and structured output. For a gateway pilot, start with its advertised image/text route; test a chart or screenshot with explicit extraction rules, and validate the resulting fields against the original. Qwen specifications

Thinking effort and cost-quality tradeoffs

The native API offers low, high and xhigh effort, defaulting to high. Run a controlled pilot: keep inputs and grading fixed, measure accepted results and billed output, then increase effort only where it improves the task. An effort name is not a cross-vendor compute standard.

WorkflowAcceptance criteria
Engineering and codingReproducible steps, passing checks and explicit assumptions.
Research and document workCitations support each material claim; conflicts stay visible.
Long-running agentsReviewable checkpoints, bounded spend and recoverable failures.

Qwen4-Max official evaluations and real cases

Official demonstrations: deliverables, not just answers

Official documentation describes an Unreal Engine game, financial research with linked spreadsheets and presentations, and legal due diligence with traceable sources. These are vendor demonstrations with supporting tools, not fresh QVeris measurements. Official case summary

The research examples distinguish experimentally tested cooling-fluid work from computational drug-candidate screening that still requires laboratory validation. In your own pilot, preserve that boundary between a proposed result and a verified result.

How to read benchmark claims

Check the release report for model snapshots, effort settings, available tools and grading conditions before comparing scores. This page does not reproduce unverified benchmark numbers or combine incompatible leaderboards. Qwen4-Max release report

A reproducible five-task pilot

Test questionWhat to check
Investment researchRevenue 550.8, EBIT 112.084, FCFF 68.063; traceable formulas and conflicting assumptions.
Acquisition diligenceClause-level citations, unsigned changes excluded, unresolved consent and data risks tracked.
Production code repairAtomic updates, idempotency, stale events and rollback tests; external side effects handled separately.
3D product creationRunnable parameterized script, hollow pipes, isolated reruns and reversible exploded view.
Scientific validationNo candidate passes all gates; control operating conditions before comparing thermal results.

These original prompts adapt task types described in Qwen’s professional-work and scientific cases, Kimi’s engineering and 3D demonstrations, and DeepSeek’s project-development examples. Inputs are fictional, self-contained evaluation fixtures. They are not official benchmark questions or scores. Code and 3D tasks require a separate execution environment to validate artifacts; a chat answer alone does not establish a passed test. Qwen cases · Kimi examples · DeepSeek examples

Qwen4-Max usage guide and FAQ

Long output and context limits

A million-token context is not a million-token output allowance. Native documentation lists up to 262,144 tokens for thinking plus the final answer, or 131,072 without thinking. Reserve room for generation and tool messages; start with a realistic output budget.

Is Qwen4-Max the same as the Preview model?

No. Use qwen4-max for this gateway route. Do not carry over Preview allowances, limits or subscription terms.

Can I use an existing OpenAI SDK integration?

Configure the QVeris base URL, use a QVeris key and select the current public model ID. Then verify tools, structured output and streaming against your application requirements.

Does gateway access include live data and display rights?

Model access, live data access and data rights are separate. Choose a data provider for freshness and coverage, then check its terms for commercial display or redistribution. A model response alone is not a live data feed.

How should I pick between Qwen4-Max and other models?

Use the same task set, run several trials and compare accepted deliverables per total spend. Include failed attempts and retries. Token price alone cannot establish the cheapest completed task.

Official documentation and related reading