T
TENCENT / tencent/WeVisForensics

WeVisForensics-4B

A forensic system with five LoRA adapters on Qwen3.5-4B. An investigator coordinates appearance, physics, common-sense and visible-text inspectors alongside speech tools. The BF16 adapters were published on November 19.

01 / SUMMARY

The essential

ORIENTATIONAI-generated video detection

Primary use stated or inferred cautiously from official documentation.

CONTEXTVideo-window inspection

Maximum input capacity when published by the source.

EXITVerdict and explanation; does not generate video

Documented maximum generation limit.

ACCESSHugging Face · downloadable checkpoint

Verified availability channels.

MODEL TYPEMultimodal · Video

Primary functional family and verified specialties.

LICENSEApache 2.0

Declared terms for API access, model weights, or self-hosting.

02 / TECHNICAL SHEET

Limits and integration

API IDtencent/WeVisForensics
Model typeMultimodal · Video
Access modelOpen weights
LicenseApache 2.0
DeploymentSelf-hosted
Release19 NOV 2026
Knowledge cutoffNot published
EntranceVideo · Audio · Text
ExitText
Context windowVideo-window inspection
maximum outputVerdict and explanation; does not generate video
ReasoningNot published
Published tools
Structured OutingsNot published
Batch processingNot published
Prompt cacheNot published
Fine-tuningNot published
Verified platformsHugging Face · Transformers
BEST SUITED FOR

Research into audiovisual forgery and explainable analysis of indicators.

WORTH MONITORING

These are adapters, not a standalone model: the base and additional tools with their licenses are required. False positives are possible; it does not prove authorship and must not make autonomous decisions about people. The single-GPU recipe specifies at least 24 GB.

03 / CAPABILITIES

What it can do

01

Orientation

Research into audiovisual forgery and explainable analysis of indicators.

02

Context

Video-window inspection · Verdict and explanation; does not generate video

03

Tools and integration

Not published

04

Access

Hugging Face · downloadable checkpoint

Input modalities
VideoAudioText
04 / PRICES

Documented cost

ConceptWorthUnit/condition
Standard inputNo published official API rateStandard API rate
Cached inputNot publishedReading reused prefixes
Cache write or storageNot publishedThe condition varies by provider
Standard outputSelf-hosted infrastructure costMay include reasoning tokens
Batch inputNot publishedAsynchronous processing
Batch outputNot publishedAsynchronous processing

No verified official API tariff. Self-hosting requires memory, compute and maintenance; downloadable weights do not make inference free.

i Prices change and may depend on level, region or context length. Check the source before making a decision.

05 / EVALUATIONS

How to read the results

“

The DTPO system reports 97.7% accuracy on GenBuster-Bench and 97.5% out-of-domain on GenBuster-Bench++. These concern the complete system with inspectors and tools, not an isolated adapter.

BenchmarkResultMetricSource
GenBuster-Bench97.7 %AccuracyView source ↗
GenBuster-Bench++ (out-of-domain)97.5 %AccuracyView source ↗

Inferama only highlights a “best result” when the metric, test set, configuration, and date allow for an equivalent comparison. The supplier's figures are presented as claims from its own source.

06 / VERSIONS

Chronology

tencent/WeVisForensics

BF16 checkpoint.

07 / SOURCES

Traceability