Falconsai NSFW Model: Cost & Download Reality (2026)
A 2025 Falconsai NSFW classifier story resurfaced. We verify current downloads, model changes, license, hosting costs, and deployment risks.
By AI Pricing Guru Editorial Team
AI Pricing Guru articles are maintained by the editorial workflow behind the site: daily pricing snapshots, provider source checks, and review passes for model launches, subscription limits, and billing changes.
TL;DR
- The alert recirculates a February 2025 article; this is not a new model launch or current top-six ranking.
- Hugging Face currently shows about 4.1 million recent-window downloads and 1.38 billion all-time, but its counters measure file requests—not unique users.
- The core ViT weights date to 2023; the September 7, 2026 repository change edited only the README.
- The Apache-2.0 model has no per-token rate: deployment cost comes from compute, uptime, scaling, monitoring, and moderation review.
- Falcons.ai separately publishes API plans from $9.99/month and a $5,000/month self-hosted container; it does not document those routes as this exact checkpoint.
A February 14, 2025 story about Falconsai/nsfw_image_detection resurfaced on Hacker News on September 8, 2026. It reported 50.8 million downloads in 28 days and a sixth-place Hugging Face ranking at that time.
That snapshot should not be presented as breaking news today. The official repository shows that the Vision Transformer model was created in 2023, a quantized YOLO file was added in April 2025, and the latest September 7 change updated only the README. No new weights, benchmark, hosted price, or commercial API launched this week.
The repository belongs to Falconsai and classifies images as normal or nsfw. It is not a new release in the Technology Innovation Institute’s Falcon language-model family.
What actually changed
The fresh event is a model-card edit. Hugging Face dates the ViT model.safetensors file to November 2023 and the quantized YOLO artifact to April 2025. The September 7 commit added Model Surgeon information to README.md.
The model remains an 85.8-million-parameter ViT classifier under Apache 2.0. Its card says it was fine-tuned on roughly 80,000 proprietary images split into two classes.
The 2025 claim versus current evidence
| Check | February 2025 story | Verified September 8, 2026 | What it means |
|---|---|---|---|
| Publication status | New article | Nineteen-month-old article resurfaced | Not a current launch |
| Download counter | 50.8M over 28 days | About 4.1M in Hugging Face’s current recent window | Old popularity claim is not today’s rate |
| Global rank | Sixth in the article’s snapshot | No current sixth-place evidence supplied | Do not repeat the old rank as current |
| Model weights | Existing ViT classifier | Same 2023 ViT weights | No model upgrade this week |
| Latest repository activity | Not applicable | README-only edit on September 7 | Documentation change, not new inference capability |
| License | Apache 2.0 | Apache 2.0 | No model-license fee, subject to the license terms |
Hugging Face’s current page also exposes an all-time counter above 1.38 billion. That does not mean 1.38 billion people downloaded the weights. Hugging Face documents that its model count is based on HTTP GET and HEAD requests to selected repository files; it does not deduplicate unique users and can include automated systems.
Pricing impact: open weights are not free moderation
There is no token rate to add to our canonical dataset. The classifier can run locally, through Hugging Face’s shared service, or on a dedicated endpoint. Dedicated endpoints bill for compute time; self-hosting moves the bill into hardware or rental, idle capacity, monitoring, updates, and engineering.
| Deployment route | Billing basis | Main cost risk | Best fit |
|---|---|---|---|
| Local or on-premises | Owned compute and operations | Low utilization or unsupported hardware | Private, steady workloads |
| Shared hosted inference | Provider service limits and terms | Quotas, cold starts, and changing availability | Evaluation and low-volume use |
| Dedicated endpoint | Running compute and replicas | Paying for idle capacity | Predictable production traffic |
| Rented GPU or CPU | Instance time plus operations | Oversizing and weak autoscaling | Teams needing infrastructure control |
Falcons.ai also operates a separate authenticated visual-safety API. Its live billing configuration currently publishes these plans:
| Falcons.ai plan | Monthly price | Published allowance |
|---|---|---|
| Free | $0 | 100 successful requests per model |
| Plus | $9.99 | 3,000 successful requests shared across models |
| Pro | $29.99 | 10,000 shared successful requests |
| Ultra | $299.99 | 100,000 shared successful requests |
| Containerized Self-Hosting | $5,000 | Unlimited calls inside the buyer’s deployment |
At full pool use, Plus works out to about $0.00333 per successful request; Pro and Ultra work out to about $0.003. These are utilization calculations, not separate per-image list prices. The vendor does not publicly establish that its paid API or enterprise container serves the exact Falconsai/nsfw_image_detection repository revision.
Use the maintained Falcons.ai pricing page for the current plan data and source boundary.
Do not put this image classifier into a token spreadsheet. The AI token calculator applies to token-priced models, while Together AI pricing and Mistral pricing cover different, generative-model buying decisions. Our local AI versus API guide explains the shared utilization tradeoff.
Accuracy claims need a production test
The model card reports 98.04% evaluation accuracy, but it does not publish the proprietary dataset, class balance, subgroup analysis, threshold calibration, false-positive rate, or an independent benchmark. A single accuracy number is not enough for a moderation policy.
Binary normal versus nsfw labels are coarse. Production systems may need separate handling for sexual, suggestive, medical, artistic, violent, and age-uncertain content. False positives suppress legitimate content; false negatives create safety and compliance risk.
Who benefits—and who does not
Teams needing an inspectable moderation component benefit from the permissive license, Transformers integration, and local option. Privacy-sensitive applications can keep images within their boundary, while high-volume services can test whether it reduces calls to a larger moderation stack.
The download headline benefits Falconsai as a distribution signal, but it does not prove accuracy, active installations, or commercial revenue. Platforms needing audited policies, appeals, multilingual context, age estimation, or nuanced categories should not treat a popular binary classifier as a complete trust-and-safety system.
What deployment teams should do
- Pin an exact repository revision; do not deploy a moving
mainbranch without review. - Test representative images across languages, regions, skin tones, medical contexts, art, and borderline cases.
- Measure precision and recall at the threshold that matches your harm policy—not only aggregate accuracy.
- Send uncertain cases to a second model or human review, and build an appeals path for consequential decisions.
- Compare CPU and GPU throughput at expected concurrency before reserving dedicated capacity.
- Track cost per reviewed image, including false positives, manual review, storage, and incident handling.
For rented infrastructure, compare Runpod only after benchmarking the classifier on your own image mix.
Affiliate disclosure: AI Pricing Guru may earn a commission from the sponsored Runpod link at no extra cost to you. It does not affect this analysis.
Bottom line
Falconsai’s NSFW classifier remains a heavily downloaded open model, but today’s alert is a recycled 2025 popularity story—not a new launch, price change, or verified current ranking. The September 2026 repository activity was documentation-only.
The model is outside the scope of the text-based AI Pricing Guru Labs leaderboard, and there is no checkpoint-specific public unit price to score. We recorded an explicit image-harness blocker. Buyers should evaluate it as a self-hosted, compute-billed, or separately subscribed moderation component, with cost measured per correctly reviewed image.
Sources: the resurfaced Middle East AI News article; the official Falconsai model card, repository history, and Hugging Face download-count methodology; Hugging Face endpoint pricing; and Falcons.ai’s live public billing configuration. Facts and prices checked September 8, 2026 at 14:05 UTC.