All stories
VIGOR-27B Launches Community reports

VIGOR-27B offers a strict, downloadable image moderator

The independent 27B release targets Safe/NSFW decisions with scene descriptions and JSON output. Its conservative policy deliberately flags ambiguous imagery.

By AITrending2 min read
Sources (4)

Why now VIGOR-27B weights were uploaded on November 1 at 12:10:13 UTC, within the current 48-hour window. November 2 model-card edits explain the release but are not a new checkpoint launch.
VIGOR 27b announcement banner from prithivMLmods, labeled Safety Classification.
Author-published VIGOR-27B announcement artwork from the English model card. This is release branding, not a classification example or evidence of model accuracy. Source: prithivMLmods

VIGOR-27B gives developers downloadable weights for an unusually strict image-moderation workflow. Independent developer prithivMLmods describes the dense 27B multimodal model as based on Qwen3.8-27B and designed to return Safe/NSFW classifications alongside a scene description and explanation.

The weight upload is dated November 1 at 12:10:13 UTC. The English model card received further edits on November 2; those edits should not be confused with a new weight release or an official Qwen announcement.

A moderation policy developers can inspect

The card’s standard prompt sets a conservative threshold: nudity, revealing underwear, suggestive swimwear, see-through clothing and sexualized illustrations or poses are to be flagged. Its instruction for ambiguous cases is explicit: classify them as NSFW.

That makes the policy itself a central integration decision. A service applying different standards to fashion, art or ordinary beach photography would need to evaluate whether this threshold fits its audience. The card acknowledges that benign content may be flagged, unsafe content may be missed, and results depend on image quality, context and instructions.

The requested JSON fields are scene_description, is_nsfw, classification_reason, nsfw and safe. The last two are complementary integer flags. This provides a documented format for connecting a classifier to review queues, but the model card’s examples do not establish that every generated response will meet the schema.

What is available to run

The repository supplies two Safetensors weight shards, processor and tokenizer files, a chat template and configuration. Its weight index records approximately 55.56 GB of tensor data, before runtime overhead; that is not a tested accelerator-memory requirement.

The English quick start uses Transformers, Accelerate and PyTorch with automatic device placement. The author says a simple classification request can be used in Transformers, while the longer standard prompt is intended for GGUF quantizations and other implementations. That documentation does not mean this repository includes GGUF files. The card labels the release Apache 2.0, subject to applicable base and component terms.

The author also claims classification improvements over an earlier VisionGuard variant and faster speculative decoding from preserved multi-token prediction tensors. The reviewed card provides no benchmark table or measured timings establishing those gains. AITrending inspected the English documentation and release files but did not run the model. The concrete news is a downloadable moderation model with an explicit, stringent policy that developers can now evaluate.

Related reading