DeepSeek V4 Pro vs GLM-5.3: Which is the top domestic large model kit?

Today I'm posting some information related to adversarial testing of the new model.

DomesticLarge ModelTwo powerful new releases have recently emerged: DeepSeek-V4-Pro and GLM-5.3.

After the two models were released, online reviews were inconsistent, and it was difficult to judge the real difference based on benchmark scores alone.

So this time we won't discuss parameters, let's go straight to the task and use the same...Prompt wordsComplete each item with the same materials and the same delivery requirements.

Both models are in Claude The code runs within the same environment, with identical permissions and initial files; each case runs only once, without any adjustments made for individual models.Prompt words。

After the run, let's see who is the top male model in China and who is the top male model in China.

Case 1 Accident Evidence Collection and Claim Calculation

Prompt words：

Based on the following mock contract, monitoring records, work orders, status pages, and bills, determine whether the service in August 2026 met the SLA and prepare the claim materials.

All times are Beijing time UTC+8.

Document 1: Article 8.2 of the Main Service Agreement

The monthly service fee is 200,000 yuan. Availability credit is calculated based on the monthly service fee.

– Monthly availability is below 99.90% but not below 99.50%, with a credit limit of 5%.

– Monthly availability is below 99.50% but not below 99.00%, with a credit limit of 10%.

– Monthly availability is below 99.00%, and credit limit is 20%.

– The maximum credit limit for the month is 20% of the monthly service fee.

Document 2: Article 3 of the SLA Annex

– Monthly availability = 1 – Number of unavailable minutes in the current month ÷ Total number of minutes in the current month.

– The service is considered unavailable when more than 20% of active tenants fail the core health check.

– Supplier monitoring records are the primary evidence for availability calculations.

Status pages and customer service emails can only be used as supplementary evidence.

– Approved scheduled maintenance can be excluded, but the supplier must notify the customer at least 5 full business days in advance.

– Up to 4 hours of scheduled maintenance can be excluded per month.

Document 3: Article 4 of the SLA Annex

P1 incident timeframe is calculated from the earlier of the customer report or the first alarm detected by the supplier:

– Response time: 15 minutes.

– Temporary solution duration: 4 hours.

Service restoration time: 8 hours.

Document 4: Supplier Monitoring Export

2026-08-05 01:00 to 02:30:

The percentage of tenants who failed the core health check was 100%.

2026-08-12 09:12 to 15:42:

The percentage of tenants who failed the core health check was 100%.

2026-08-22 02:00 to 03:10:

The percentage of tenants who failed core health checks was 35%.

The three time periods do not overlap.

Document 5: Planned Maintenance Notice

Notification release time: 2026-08-01 18:00.

Scheduled maintenance time: 2026-08-05 01:00 to 02:30.

The notice states that core services may be unavailable during maintenance.

Document 6: Work order dated August 12

First alert from supplier monitoring: 09:12.

Customer created P1 work order: 09:16.

Supplier's first manual response: 09:38.

Temporary solution complete: 12:50.

Supplier status page flag restored: 15:10.

The monitoring system has returned to normal: 15:42.

Work order closed: 17:30.

Document 7: Account Manager Email

"The incident on August 12 was fully resolved at 15:10, so the recovery time did not exceed 6 hours."

Document 8: August Bill

Monthly service fee: 200,000 yuan.

One-time implementation service fee: 50,000 yuan.

Taxes and fees are not included in the calculation of SLA credit limits.

Please deliver:

1. Evidence sheet, listing each fact, source, level of evidence, and conflict.

2. Determine whether the three unavailable times in each segment should be included in the SLA.

3. Calculate the total number of minutes, unavailable minutes, and monthly availability for August.

4. Determine the corresponding credit limit percentage and amount.

5. Review the response, temporary solutions, and recovery timeline from August 12.

6. Explain whether the conclusions in the account manager's email are valid.

7. List the items for which there is insufficient information or which cannot be determined.

8. An incident summary submitted to management.

9. A formal claim letter sent to the supplier.

10. An incident timeline suitable for inclusion in a management presentation.

11. Check all time differences, proportions, and amount calculations.

DeepSeek-V4-Pro

GLM-5.3

Both models calculated a credit limit of 40,000 yuan and also identified the errors in the supplier's email.

DeepSeek-V4-Pro is more like a contract reviewer, breaking down 23 pieces of evidence in detail, clearly explaining which ones are valid and which ones lack evidence.

However, a minor issue with DeepSeek-V4-Pro is that it writes 98.7679% as approximately 98.76%, while the standard rounding to two decimal places should be 98.77%.

GLM-5.3 is more like a delivery manager, directly producing a claim file that can be opened, printed, and reported, with a significantly higher level of format completion than DeepSeek-V4-Pro.

Overall, GLM-5.3 won Case 1 due to its high level of delivery completion, while DeepSeek-V4-Pro demonstrated greater stability in evidence reasoning.

Case 2: Integrated Office Applications (Excel, Word, PPT)

Prompt words：

Process the following order data and generate a set of deliverable office documents.

order_id,user_id,paid_at,amount,status,channel,original_order_id

A1001,U01,2026-08-01 23:40:00,299.00,paid,web,

A1002,U02,08/02/2026 09:15,"1,299.00″,paid,ios,

A1002,U02,2026-08-02 09:15:00,1299,PAID,iOS,

A1003,,2026/08/03 14:20,-199,refund,web,A1001

A1004,U04,2026-08-03T16:30:00+08:00,399,PAID,android,

A1005,U05,2026-08-04,0,paid,web,

A1006,U06,bad-date,99,pending,Web,

A1007,U07,2026-08-04 10:15:00,699,paid,android,

R1001,U01,2026-08-05 12:20:00,-99,refund,web,A1001

R1002,U02,2026-08-05 13:30:00,-1500,refund,ios,A1002

A1008,U08,2026-08-06 15:00:00,"1,099.00″,paid,web,

A1009,U09,2026-08-06 16:00:00,499,cancelled,android,

A1010,U10,2026-08-07 11:00:00,199,paid,,

A1011,U11,2026-08-07T12:00:00+08:00,259,paid,ios,

A1012, U12, 2026-08-08 09:00:00, 299 yuan, paid, web,

Cleaning rules:

– order_id, user_id, paid_at, amount, status, and channel are all required fields.

– Dates are all in Beijing time `YYYY-MM-DD HH:mm:ss`.

– Status and channel should be in lowercase.

– The paid amount must be greater than 0.

– The refund amount must be less than 0, and a valid original paid order must be found.

– The absolute value of a single refund cannot exceed the payment amount of the corresponding order.

– Only one duplicate record with the same order_id and identical business content will be retained.

– Pending and cancelled orders may be held, but are not included in net income.

– Effective net income equals effective paid amount plus effective refund amount.

Data that cannot be reliably repaired is entered into the rejected table, retaining the original value and the reason for rejection.

Please generate:

1. `cleaned.csv`。

2. `rejected.csv`。

3. `clean_orders.py`。

4. `analysis.xlsx` contains raw data, cleaned data, rejected data, statistical summaries, and charts.

5. Excel 365 Formulas: Formulas can be recalculated as the source data changes.

6. `memo.docx`, for management to read.

7. `management_report.pptx`, 7 pages in total.

8. A verification report listing the original number of rows, the number of valid rows, the number of duplicates, the number of rejections, and the net valid income.

Run the cleaning script, open or render the generated file, and inspect it. CSVCheck if the numbers in Excel, Word, and PowerPoint are consistent. Continue to correct any issues found.

DeepSeek-V4-Pro

GLM-5.3

The final business figures in DeepSeek-V4-Pro are all correct, and the verification report lists all valid orders for easy manual traceability.

However, date parsing relies solely on regular expressions, so even invalid times like 2026-13-40 25:61:00 will be accepted; when encountering a +00:00 timezone, the script simply removes the timezone marker.

Excel lacks frozen headers, resulting in excessively wide columns and a generally poor browsing experience.

The PPT is primarily text-based with ample white space.

GLM-5.3 uses Decimal to handle amounts, can correctly convert ISO time zones, and can reject invalid dates. The script uses its own directory to locate files, so it is not easy to find the wrong file when executed from other directories.

Add channel analysis, status analysis, formula charts, and frozen headers to Excel.

Word contains channel charts.

The PPT includes flowcharts, data cards, tables, and charts, and is almost ready for presentation.

DeepSeek-V4-Pro delivered a set of office-level results with accurate numbers and a basic format; GLM-5.3 was significantly more complete.

Case 3: Warehouse Concurrency Bug Fix

Prompt words：

Please handle duplicate order incidents in the Order Desk.

Warehouse path:

D:\360MoveData\Users\win\Desktop\选题\order-desk

Customer support has received multiple reports of duplicate orders.fastDouble-clicking "Save Order," retrying after a network timeout, or two requests arriving at the server almost simultaneously may generate multiple orders. Duplicate orders will proceed to subsequent processes separately and currently require manual cancellation.

Please proceed directly to the warehouse for processing:

1. Read README.md, TASK.md, application code, database migrations, and existing tests.

2. Run the existing tests to confirm the current status.

3. Reliably reproduce the issue of duplicate orders.

4. Trace the call process between the front-end, HTTP API, business services, and SQLite to find the specific root cause.

5. Modify the code to resolve issues related to concurrent requests, network retries, and duplicate orders after service restarts.

6. Maintain the existing API and normal order placement, order query, and list functions.

7. Supplement tests related to concurrency, duplicate requests, conflicting requests, transaction failure retries, and service restarts.

8. Examine the impact of the database migration on the new and existing databases.

9. Run all tests, and if errors are found, continue to fix them and re-verify.

Please make the changes directly in the repository; do not just provide suggestions.

Instructions upon completion:

– Root cause.

– Remedial measures.

– Modify the file.

– Database migration.

– Added test.

– The actual commands and results executed.

– Risks still exist.

DeepSeek-V4-Pro

GLM-5.3

ON CONFLICT(customer_id, idempotency_key) DO NOTHING

Case 4: File Export Security Hardening

Prompt words：

Review and fix the following FastAPI interfaces:

from fastapi import FastAPI

from fastapi.responses import FileResponse

from pydantic import BaseModel

import os

import requests

app = FastAPI()

class ExportRequest(BaseModel):

url: str

filename: str

@app.post("/export")

def export_pdf(req: ExportRequest):

output = f"/tmp/{req.filename}"

html = requests.get(req.url).text

open("/tmp/page.html", "w").write(html)

os.system(f"wkhtmltopdf /tmp/page.html {output}")

return FileResponse(output)

Complete the following tasks:

1. List the vulnerabilities, exploitation conditions, and impacts in order of severity.

2. Provide runnable attack or test cases for the main vulnerabilities.

3. Submit the complete, fixed code.

4. Add pytest test.

5. Limit the protocol, port, target host, and number of redirects.

6. RefusalLoopPrivate, link-local, reserved addresses, and cloud service metadata addresses.

7. Simultaneously check IPv4, IPv6, IPv4 mapped to IPv6, decimal IP, and mixed-encoded addresses.

8. Handle DNS rebinding and changes in resolution results.

9. Recheck the target after each redirection.

10. When establishing a connection, fix the verified target IP.

11. Limit response header size, response body size, connection time, read time, and total time.

12. When using streaming, the entire response cannot be loaded into memory before the size is verified.

13. Prevent command injection, path traversal, symbolic link attacks, and arbitrary file overwriting.

14. Disable shell access and restrict local file access and external resource loading when calling wkhtmltopdf.

15. Each request uses a separate temporary directory.

16. Input and output files cannot be shared between concurrent requests.

17. Temporary files should be cleaned up when the client disconnects, the request is canceled, the conversion fails, and the transmission is completed.

18. Handling cleanup timing while FileResponse is still reading a file.

19. Set timeouts, resource limits, and termination procedures for the conversion process.

20. Run tests and continue to fix any issues found.

The test should at least cover:

– `127.0.0.1`

– `::1`

– `169.254.169.254`

– Private IPv4

– IPv4 mapping to IPv6

– DNS rebinding

– Redirect toPrivate address

– Ultra-large response

Slow response

– Command injection filename

– Path crossing filename

– Symbolic link output

– Two concurrent export requests

– Client canceled midway

wkhtmltopdf timed out

Deliver vulnerability reports, fixes, test code, run instructions, and the remaining boundaries.

DeepSeek-V4-Pro

GLM-5.3

Case 5: External sorting of large files

Prompt words：

Implement the command-line tool `ext-sort` using Go.

The input is a CSV file with a maximum size of 10GB, and the maximum available memory is 512MB.

Sorting rules:

1. `created_at` in ascending order.

2. If the times are the same, sort by `user_id` in ascending order.

3. If the first two items are the same, sort them in ascending order of their original row numbers to ensure a stable sorting.

Require:

– Correctly handle quotes, commas, UTF-8 encoding, and line breaks within fields.

– Keep the table header.

– created_at accepts RFC3339 and `YYYY-MM-DD HH:mm:ss`.

- Records that cannot be parsed are written to `rejected.csv`.

A single bad data entry cannot cause the entire task to fail.

– Supports `--input`, `--output`, `--temp-dir`, `--memory-limit`, and `--workers`.

- The entire file cannot be read into memory.

– Memory limits should include the main usage of sorting blocks, buffers, and merge phases.

- Clean up temporary files when a termination signal is received.

– Provides a clear error message when there is insufficient disk space, file corruption, or insufficient permissions.

– The target output file can only be replaced after the entire generation is complete.

– The sorting results are reproducible.

- Multiple runs should not leave behind unrecognizable temporary files.

Please deliver:

1. A complete Go project.

2. Unit testing, integration testing, and property testing.

3. Random test data generator.

4. Sorting result verification tool.

5. Performance testing.

6. Crash recovery and temporary file design specifications.

7. Build and run commands.

A test file larger than the available memory was actually generated, and the process was run to sort, verify the results, and measure peak memory usage. Errors were corrected as they were discovered.

DeepSeek-V4-Pro

GLM-5.3

Case 6 High-Density Data Dashboard

Prompt words：

Creating "Using React, TypeScript, and Vite"AI The "Model Evaluation Lab" single-page application.

Visual direction:

– Dark graphite background.

Acidic green is only used for selected states, error messages, and a small amount of important data.

– Gradients, emojis, glass-like effects, and full-screen rounded corner cards are prohibited.

- High information density, maintaining a clear visual hierarchy.

– Establish a stable system of spacing, fonts, and colors.

– Avoid layouts and decorations common in backend templates.

Data scale:

– Generate 50,000 evaluation run records locally.

Each record includes the model, evaluation set, time, accuracy, latency, cost, input token, output token, and running status.

– Data generation uses a fixed random seed.

Page content:

– Model and evaluation set selection.

– Run parameters panel.

– Comprehensive indicator matrix.

– Delay distribution map.

– Cost trend chart.

– Graph showing the relationship between accuracy and token consumption.

– A virtual scrollable running history table.

– Single run details drawer.

– Multi-model comparison mode.

Interaction requirements:

– Filter criteria are synchronized to the URL.

– The browser's forward and back navigation correctly restores the filtering state.

– fastWhen continuously switching filter criteria, the results of older requests cannot overwrite the results of newer requests.

– Supports loading, empty data, error, request cancellation, and partial failure states.

– The chart displays tooltips, legends, outliers, and clear coordinate units.

– Coordinate axes cannot be misleading by being truncated.

The table supports sorting, filtering, pagination, or virtual scrolling.

– All interactions support the keyboard.

– Focus status is clear.

– Supports `prefers-reduced-motion`.

- Compatible with 1440, 768 and 390 widths.

– Use the local mock API to support delayed, error, and canceled requests.

– Add component testing, state testing, and key interaction testing.

After completion:

1. Run format checks, type checks, tests, and production builds.

2. Launch the application and check the console.

3. Measure the initial rendering and filtering response with 50,000 data points.

4. Generate screenshots with widths of 1440, 768, and 390 respectively.

5. Check for overflow, occlusion, alignment, text readability, and misleading charts.

6. Use the keyboard to complete a full filtering and details viewing process.

7. Fix the identified issues and re-verify.

DeepSeek-V4-Pro

GLM-5.3

DeepSeek-V4-Pro's pages are more comfortable, especially at widths of 390px and 768px, where filter areas, indicator cards, and charts remain clear.

GLM-5.3 provides more comprehensive coverage of the original requirements. With a fixed seed and a fixed timeline, the generated data remains consistent regardless of the machine or date of execution. The charts also provide a data table view, with initial data rendering taking approximately 359ms and filtering response time around 286ms, slightly faster than DeepSeek-V4-Pro.

GLM-5.3 has a slight lead in status handling and acceptance, while DeepSeek-V4-Pro performs better in page readability and multi-model analysis.

Case 7 Node Editor

Prompt words：

Implement a lightweight workflow editor using React, TypeScript, and SVG.

Functional requirements:

– Create, select, drag, rename, and delete nodes.

– Create a connection by dragging a cable from the port.

- Reject self-loops, duplicate connections, and illegal directions.

– Single selection, multiple selection, and box selection.

– Copy, paste, and duplicate nodes.

– Canvas zoom and pan.

– NodeautomaticEnter the visible area.

– Undo and redo, keeping at least 50 steps.

– Supports Delete, Backspace, Ctrl/Cmd+C, Ctrl/Cmd+V, Ctrl/Cmd+Z and Ctrl/Cmd+Shift+Z.

– When editing the input box, shortcut keys should not be used to accidentally delete nodes.

– JSON import and export.

– JSON errors require specifying the exact field or location.

- Restore from localStorage after refresh.

It can still operate with 500 nodes and 1000 connections.

– Provides a small map or canvas positioning tool.

– Supports keyboard selection of nodes.

Project requirements:

1. First, design the state model, coordinate system, and historical record structure.

2. Business status and temporary drag-and-drop status are managed separately.

3. Undo records cannot save meaningless intermediate mouse movement states.

4. The copied node ID, location, and connections must be updated correctly.

5. Even after scaling, dragging, selecting, and port connections must remain accurate.

6. Write tests for connection rules, undo/redo, copy and paste, and import/export.

7. Add performance testing for large data volumes.

8. Run type checks, tests, and production builds.

9. Launch the application to check the interaction and fix issues.

Deliver a complete project, state design specifications, operational methods, and actual test results.

DeepSeek-V4-Pro

GLM-5.3

DeepSeek-V4-Pro's strengths lie in its architecture and page completion. It features visual consistency across the dark canvas, ports, connections, and minimap.

The advantage of GLM-5.3 is its more robust coverage of requirements. Start, Step, and End nodes have clearly defined color and port rules; repeated pasting continuously increases the offset; the keyboard arrow keys can be used to select nodes by spatial location; and when selecting nodes outside the canvas, further options are available.automaticAdjust the viewport.

GLM-5.3 has more complete features and stronger test evidence; DeepSeek-V4-Pro has a better drag-and-drop design.

Case 8 3D Product Configurator

Prompt words：

Create a modular 3D configurator for a table lamp using Three.js, TypeScript, and Vite.

Modeling requirements:

– Use procedural geometry to create the lamp holder, two lamp arms, pivot, lampshade, and bulb.

– Each component is an independent object.

– The components are connected in a reasonable manner and should not detach when rotating the lamp arm.

– Downloading external models and textures is prohibited.

Interaction requirements:

– Clicking on a component will display its outline and name.

– Supports three materials, four colors, and three light levels.

– Supports adjusting the angles of the two lamp arms and lampshade.

– Each joint has a reasonable range of rotation.

– Supports exploded view animation and dimension annotation.

– Provides a configuration summary, price changes, and a reset button.

– Supports mouse, touch, and keyboard.

Rendering requirements:

– Use PBR material.

– It has ambient light, key light, and soft shadows.

– The light bulb produces observable changes in illumination.

– OrbitControls has reasonable target, distance, and polar angle limits.

– The maximum DPR is 2.

– Correctly adjust the canvas and camera after the window changes.

– Pause unnecessary rendering when the page is not visible.

– Display a downgrade prompt when WebGL is unavailable.

– Properly release geometry, materials, and event listeners.

After completion:

1. Run type checks, tests, and production builds.

2. Launch the page and check the console.

3. Check for joint dislocations, clipping, shadows, Z-fighting, and lens cropping.

4. Check mobile device operation.

5. Generate screenshots of the default, exploded, lit, and mobile views.

6. Continue to make corrections based on the actual footage.

7. Submit the complete project, how it runs, and any known limitations.

DeepSeek-V4-Pro

GLM-5.3

DeepSeek-V4-Pro has an advantage in 3D modeling. The proportions of the lamp base, two lamp arms, joints, and lampshade are well-proportioned, making it immediately recognizable as an adjustable desk lamp; the nested joints are also correctly implemented, so when the parent lamp arm is adjusted, the subsequent parts move accordingly.

GLM-5.3 is more like a product configurator that allows for direct demonstration. Materials, colors, lighting, angles, real-time prices, and price differences are all grouped into a single card, making the information easier to find than in DeepSeek-V4-Pro; the model is broken down into eight optional parts, and the bulb point light source even has shadows enabled.

The main problem with GLM-5.3 lies in the visual results. The lampshade is too large, the default lens is facing the inside of the lampshade, and when the two lamp arms are covered, it looks more like a ring light.

DeepSeek-V4-Pro offers better 3D modeling, selection effects, and mobile readability; GLM-5.3 leads in configuration panel, but its obvious visual issues significantly impact 3D case studies.

Case 9: Procedural Modeling

Prompt words：

Write a Python script for Blender 4.x to generate a coffee kiosk.

Default size:

– 4 meters wide.

– 3 meters deep.

– 2.8 meters high.

– The counter is 1.05 meters high.

Model content:

– Platform, walls, ceiling, and service windows.

Counters, lockers, and shelves.

– Coffee machine, grinder, cups, and menu board.

– External signage.

– Power outlets and cable trays.

– Camera, sunlight, and two area lights.

Structural requirements:

– Architecture, Furniture, Props, and Lights are divided into independent collections.

– Use stable and clear English names for the main objects.

– The width, depth, counter height, and number of shelves of the kiosk can be modified via parameters.

– After the parameters change, the position and proportion of each component are adjusted synchronously.

– Add appropriate chamfers and normals.

– Use procedural materials.

– The main dimensions are written to the object's custom properties.

- When updating a collection managed by a script during repeated runtime, duplicate objects must not be generated.

– Do not delete user objects from other collections.

deliver:

1. A complete Python script.

2. Generate and save `.blend`.

3. Render three perspectives: front view, 45-degree angle, and interior work area.

4. Check for gaps, overlaps, normals, material proportions, and exposure.

5. Rerun the program after modifying the dimensions and the number of shelves.

6. Ensure that object names and custom properties remain consistent.

7. Provide the running method, parameter description, and object structure description.

The script is actually executed and errors are handled in Blender 4.x.

DeepSeek-V4-Pro

GLM-5.3

The DeepSeek-V4-Pro project includes 59 objects and 14 sets of materials, categorized by architecture, furniture, props, lighting, and camera. It also provides parametric scripts, three cameras, and...automaticQuality inspection.

The DeepSeek-V4-Pro output has obvious flaws. The front sign and menu text are upside down and mirrored, and the sign is also cropped.

The GLM-5.3 has a significantly higher level of spatial completion. The front view shows the service window, shelves, cups, coffee equipment, countertops, and cabinets, and the combination of green and wood tones is closer to a deliverable commercial space plan.

GLM-5.3 also has a visible issue: the menu text does not fall within the black menu panel.

Considering the three-view drawings, spatial integrity, parameter verification, and final output, GLM-5.3 is still slightly better; DeepSeek-V4-Pro has an advantage in script structure, but incorrect text direction directly lowers the delivery quality.

Case 10 Document Review

Prompt words：

The company is preparing for final acceptance of the project and to make the final payment. However, inconsistencies have been found in the project documentation regarding the version of the requirements, the amount, the date, the responsible person, and the acceptance criteria.

Please read through all the contracts, supplementary agreements, requirements documents, meeting minutes, emails, quotations, and acceptance materials provided in this round to verify the current requirements that should be followed for the project.

Please deliver:

1. A list of requirements and commitments, each entry containing:

- serial number

– Needs, commitments, or decisions

– Current status

– Source document

– Page number, chapter, or paragraph position

– Original evidence

- Effective Date

– Amount Involved

– Person in charge

– Acceptance conditions

– Has it been modified or repealed by subsequent documents?

– Basis for modification or repeal

– Issues requiring confirmation

2. List of file conflicts, focus on checking these:

– Contracts and Supplementary Agreements

– Different versions of the requirements document

Meeting minutes and confirmation emails

– Text and Tables

– Quote, budget and payment amount

– Planned date vs. actual date

– Person in charge and inspector

3. Old requirements that have been modified or repealed but are still cited in subsequent materials.

4. Discrepancies between contractual commitments, current implementation results, and acceptance materials.

5. Claims lacking formal evidence to support them.

6. High-risk matters that may affect acceptance, payment, or liability determination.

7. Unable to read, page missing, garbled text, or incomplete file content.

8. Issues requiring confirmation from the project manager, legal department, finance department, or supplier.

Processing requirements:

– All conclusions must include the specific file location.

– When you cannot find evidence, write “not found” instead of filling in the blanks based on common sense.

– If multiple versions of the same item exist, they will be listed separately; you cannot choose one of them yourself.

– Only when subsequent documents explicitly modify the old requirements can it be determined that the old requirements have become invalid.

– Distinguish between formal contracts, signed documents, confirmation emails, meeting discussions, and personal opinions.

– For individuals with the same name, their identity needs to be verified by combining their department, email address, or the context.

– The date, amount, responsible person, and acceptance conditions are checked item by item.

– When discrepancies are found between the text and the table, retain both pieces of evidence.

Finally, an acceptance review opinion is provided for management to read, explaining whether acceptance is currently possible, whether the final payment can be made, matters that need to be temporarily suspended, and the evidence for each judgment.

DeepSeek-V4-Pro

GLM-5.3

DeepSeek-V4-Pro accurately identified all the facts that would change the acceptance and payment conclusions, and also identified 42,000 yuan of duplicate training fees and 25,200 yuan of invalid tax, and correctly handled version conflicts; the main oversight was in the file retention requirements.

GLM-5.3 also yielded correct conclusions regarding acceptance, payment, and amount, and the evidence ledger was more detailed.

However, GLM-5.3 suffers from data errors due to over-analysis, while DeepSeek-V4-Pro is more reliable in terms of factual accuracy required for high-risk reviews.

Case 11 Customer Service Ticket Processing

Prompt words：

The customer service system needs to convert new work orders into JSON that can be read by downstream dispatch services.

The fixed structure accepted by downstream services is as follows:

{

"tickets": [

{

"id": "string",

"category": "billing|bug|account|feature",

"priority": "p0|p1|p2|p3",

"requires_human": true,

"reason": "string"

}

],

"summary": {

"total": 0,

"p0_count": 0

}

}

Classification rules:

– Account theft, unauthorized charges, or continued charges after subscription cancellation are categorized as p0.

– Software crashes that can be reproduced reliably are classified as p1.

- Unable to log in andautomaticRecovery method failed, classified as p1.

– Common functional abnormality, classified as p2.

– Feature suggestion, categorized as p3.

– Issues involving funds, account security, or identity verification require manual processing.

– Fields not defined by downstream services cannot be added to JSON.

Pending work orders:

T01: My phone was lost last night, and three unknown charges were made early this morning.

T02: The application crashes every time I try to export a PDF, and the problem persists even after reinstalling.

T03: We hope to add a dark mode.

T04: Forgot password? I haven't received the reset email, and I've checked the spam folder.

T05: I canceled my subscription two months ago, but I've still been charged for the past two months.

T06: Occasionally, a blank screen appears after filtering orders; refreshing the page will restore it.

Please return valid JSON that can be directly parsed by downstream services; do not add Markdown code fences or other text.

DeepSeek-V4-Pro

GLM-5.3

The JSON structure of DeepSeek-V4-Pro is fully compliant. The classification, priority, and manual intervention flags for all 6 work orders are correct, and the total: 6 and p0_count: 2 are also verified to be correct. The two fund security issues, T01 and T05, were judged as P0, and the account recovery issue, T04, required manual intervention, indicating that the risk boundary assessment was accurate.

The core fields of GLM-5.3 are completely identical to those of DeepSeek-V4-Pro, with no formatting errors or unnecessary explanations. GLM-5.3 retains the two-month charge from T05 in the reason section and adds that a refund will be processed manually; T04 also clearly states that manual identity verification is required, making it easier to continue processing after handing it over to customer service personnel.

The two models are not significantly different for this task.

Case 12 Project Scheduling and Cost Optimization

Prompt words：

Schedule the following software release projects and identify the lowest-cost acceleration option to meet the deadlines.

The project will commence on Monday, September 7, 2026. Only Monday through Friday will be considered; public holidays will not be taken into account.

Scheduling rules:

– Each task starts at the beginning of the workday.

– The construction period is calculated in full working days.

– Dependency tasks begin on the next business day after all prerequisite tasks are completed.

– The same person cannot participate in two tasks at the same time.

– Tasks marked with two responsible persons require both persons to be available at the same time.

Tasks that have no dependencies and have different responsible persons can be performed in parallel.

– Each acceleration option can be used at most once.

– All tasks must be completed before the end of October 13, 2026.

Task:

A. Requirements Analysis

Person in charge: Lin Zhou

Construction period: 3 working days

Dependencies: None

Cannot be accelerated

B. Interactive Prototype

Person in charge: Lin Zhou

Construction period: 4 working days

Dependency: A

It can be shortened to 3 working days, increasing costs by 12,000 yuan.

C. API Development

Person in charge: Zhou Zhe

Construction period: 6 working days

Dependency: A

It can be shortened to 4 working days, increasing costs by 28,000 yuan.

D. Data migration

Person in charge: Zhang Min

Construction period: 5 working days

Dependency: A

It can be shortened to 3 working days, increasing costs by 20,000 yuan.

E. Front-end development

Person in charge: Chen Sen

Construction period: 6 working days

Dependency: B

It can be shortened to 4 working days, increasing costs by 24,000 yuan.

F. Security Plan

Person in charge: Zhang Min

Construction period: 3 working days

Dependency: B

Cannot be accelerated

G. System integration and testing

Person in charge: Zhou Zhe, Chen Sen

Construction period: 3 working days

Dependencies: C, D, E

This can be shortened to 2 working days, increasing costs by 0.9 million yuan.

H. Security Testing

Person in charge: Zhang Min

Construction period: 4 working days

Dependencies: F, G

It can be shortened to 2 working days, increasing costs by 16,000 yuan.

I. User Acceptance

Person in charge: Lin Zhou

Construction period: 3 working days

Dependency: H

Cannot be accelerated

J. App Store Review

Person in charge: External team

Construction period: 7 working days

Dependency: I

It can be shortened to 5 working days, increasing costs by 25,000 yuan.

Please deliver:

1. The complete schedule for the non-accelerated option.

2. Completion date of the unaccelerated plan.

3. All reasonable acceleration combinations that meet the October 13 deadline.

4. Lowest cost option.

5. Item-by-item scheduling of the lowest cost solution.

6. Prove the lowest cost, explaining why cheaper combinations could not be completed on time.

7. If Zhang Min takes a day off on September 17, recalculate the lowest cost plan.

8. Schedule in Gantt chart format.

DeepSeek-V4-Pro

GLM-5.3

The core calculations of DeepSeek-V4-Pro were correct, the lowest-cost solution was found, and the completion date of October 13th was correctly calculated. Zhang Min's leave of absence on September 17th will only postpone F; the final date and the optimal solution remain unchanged.

DeepSeek-V4-Pro has omissions in combinatorial enumeration. The delivered content only includes terminal tables and ASCII Gantt charts, which are inconvenient to view and reuse.

GLM-5.3 is more complete. GLM-5.3 correctly lists all 9 non-redundant acceleration combinations, and the lowest cost, lower bound proof, and recalculation results are consistent with the independent calculation.

The generated release-schedule.html contains a summary card, three schedules, and three Gantt charts. The non-accelerated plan, the optimal plan, and the leave-request plan can be directly compared, making it easier to use as project scheduling material.

DeepSeek-V4-Pro simply calculated the solution correctly; GLM-5.3 leads in both compositional completeness and delivery format.

After running 12 cases, GLM-5.3 led in 7 cases, DeepSeek-V4-Pro led in 3 cases, and the difference was small in the other 2 cases.

GLM-5.3's advantages lie in its complete delivery. Office documents have more mature formatting, programming tasks cover more exceptions, and the front-end and node editor are closer to ready-to-use finished products. When it comes to Blender modeling and project scheduling, GLM-5.3 can deliver more complete scenes and visualization materials.

DeepSeek-V4-Pro's strengths lie in its core implementation and factual accuracy. It sorts large files faster and handles bad data more reliably; the structure, proportions, and mobile effects of the 3D desk lamp are better; and when processing high-risk acceptance data, DeepSeek-V4-Pro did not miscalculate key figures despite its pursuit of analytical depth.

If daily tasks consist of office deliverables, full projects, and long-term... Agent For execution-oriented systems, I would prioritize GLM-5.3; however, if algorithm implementation, key figures, and evidence reasoning are more important, DeepSeek-V4-Pro remains a very competitive option.

This round of top Chinese model rankings could go to GLM-5.3, while DeepSeek-V4-Pro is more like the top Chinese model.

AI The model is transitioning from the chat window to the real production process.

This round of 12 cases covers office documents, code repositories, front-end pages, 3D projects, and long document reviews. The models need to read materials, call tools, run tests, and leave deliverables that can be checked and reused. What enterprises are really willing to pay for are these capabilities that can shorten the project time and reduce rework.

The GLM-5.3 secured more wins this round, with its advantage lying in complete delivery. (Requires...)fastFor tasks involving completing solutions, prototypes, or complete sets of project materials, GLM-5.3 is easier to integrate into existing workflows.

DeepSeek-V4-Pro demonstrates greater stability in critical computations, algorithm implementation, and fact-checking. Its robust underlying capabilities are evident in large file sorting, acceptance testing, and 3D model structure. For data processing, core code, and high-risk reviews where accuracy is paramount, DeepSeek-V4-Pro remains a top choice.

A more practical business application is a task-based allocation model: GLM-5.3 is responsible for requirements coverage, project setup, and product delivery, while DeepSeek-V4-Pro is responsible for critical calculations, core implementation, and result verification.

The next round of competition for domestically produced models will also take place in these real-world projects.

Original link:DeepSeek V4 Pro (Official Version) VS GLM-5.3: Which is better?