Best Video Classification Models to Run Locally in 2026: Cosmos-Embed1-448p-anomaly-detection

Quick Answer

NVIDIA’s Cosmos-Embed1-448p-anomaly-detection is the top pick as of September 2026 because it has the newest Hugging Face publication date among the eligible candidates [12][8][1][4][10][6][2][5]. In order, the ranking is Cosmos-Embed1-448p-anomaly-detection (NVIDIA), videoprism-base-f16r288 (Google), vjepa2-vitl-fpc64-256 (facebook), vjepa2-vitg-fpc64-256 (facebook), VideoMAEv2-Base (OpenGVLab), vivit-b-16x2-kinetics400 (Google), xclip-base-patch16-zero-shot (Microsoft), and xclip-base-patch32 (Microsoft) [12][8].

Key Takeaways

• Ranking note: eligibility requires video-classification models from labs with a published paper or leaderboard record. Releases within the last twelve months are too few, so the newest available are included; every candidate has no public benchmark yet, and ordering is newest release first, then downloads.[1][2][4][5][6][8][10][12]
• NVIDIA’s Cosmos-Embed1-448p-anomaly-detection ranks first because its Hugging Face publication date, March 10, 2026, is the newest among the candidates.[12][8][1][4][10][6][2][5] Its 1.2B parameters imply 600 MB of weight memory at 4-bit, estimated (params x 0.5 bytes).[12]
• Google’s VideoPrism (videoprism-base-f16r288 (Google)) follows, with a Hugging Face publication date of June 14, 2025, and an Apache-2.0 license; parameter count, quantized memory and context are unspecified in the supplied specifications.[8]
• Established picks are facebook’s vjepa2-vitl-fpc64-256, Microsoft’s xclip-base-patch16-zero-shot and facebook’s vjepa2-vitg-fpc64-256, the three candidates with the highest downloads.[1][2][4] Their recency exemption does not change the ordering rule or grant first place.
• OpenGVLab’s VideoMAEv2-Base has 86M parameters, implying 43 MB of weight memory at 4-bit, estimated (params x 0.5 bytes); its CC-BY-NC-4.0 license restricts commercial use.[10]
• Microsoft’s xclip-base-patch16-zero-shot and Microsoft’s xclip-base-patch32 both specify a text context of 77 tokens and an MIT license; that token limit does not specify video duration.[2][5]

How do local video classification models compare on specs and benchmarks?

Which video classification models should you run locally?

1. Cosmos-Embed1-448p-anomaly-detection

Cosmos-Embed1-448p-anomaly-detection by NVIDIA ranks first because its Hugging Face publication date of March 10, 2026 makes it the newest release in this lineup.[12][8][1][4][10][6][2][5] The ordering is “newest release first, then downloads”; Cosmos has no public benchmark yet, so its position does not establish superior classification accuracy.

The model has 1.2 billion parameters and a published F32 weight size of 4.8 GB.[12] Quantized weight storage would be approximately 600 MB at 4-bit, estimated (params x 0.5 bytes) from that parameter count.[12] For local hardware planning, budget separately for runtime memory and video inputs. The weight estimate alone cannot establish a suitable GPU or total RAM requirement, and does not confirm that a working quantized implementation is available.

Video anomaly detection is the intended use indicated by this checkpoint’s name.[12] Treat it as a candidate to evaluate on your own clips; the caveat is the absence of a public benchmark for comparing its accuracy.

2. videoprism-base-f16r288

videoprism-base-f16r288 by Google ranks second under the ordering rule: newest release first, then downloads.[8][12] Its Hugging Face publication date is June 14, 2025, behind NVIDIA’s Cosmos-Embed1-448p-anomaly-detection.[8][12] The ranking includes older releases because recent candidates are sparse; Google’s entry is therefore an older fallback. Benchmark status: no public benchmark yet, so its position does not establish comparative classification accuracy.

Google distributes the checkpoint under the Apache-2.0 license.[8] The accompanying paper describes VideoPrism as a foundational visual encoder for video understanding, making that the relevant evaluation use case.[9] Parameter count, weight size and supported input dimensions are not specified here; the checkpoint name alone should not substitute for documented specifications.

For local deployment, hardware requirements remain unverified. Without a documented parameter count, a quantized weight-memory estimate would be speculative, and no GPU or RAM recommendation is justified. Evaluate the checkpoint for your video-understanding workflow, with the caveat that neither hardware fit nor classification performance is established here.

3. vjepa2-vitl-fpc64-256

vjepa2-vitl-fpc64-256 from facebook ranks third under the “newest release first, then downloads” ordering, with a Hugging Face publication date of May 31, 2025.[1] The category has too few recent releases, so older options remain eligible. An established pick, the model recorded 199,414 downloads over the preceding 30 days as of September 25, 2026; popularity does not establish classification quality.[1] Its benchmark status is “no public benchmark yet.”

For local hardware planning, its 326M parameters imply 163 MB of 4-bit weight memory, estimated (params x 0.5 bytes).[1] Published F32 weights occupy 1.3 GB.[1] The quantized figure covers weights alone; total RAM and VRAM requirements remain unestablished, so it cannot justify a specific GPU recommendation.

Choose it for local video-classification projects that require MIT-licensed weights.[1] The practical caveat is quantization: the calculated weight footprint does not confirm that a supported quantized runtime is available.

4. vjepa2-vitg-fpc64-256

vjepa2-vitg-fpc64-256 from facebook ranks fourth under “newest release first, then downloads”: its first Hugging Face publication was 2025-04-07, behind the newer entries above it.[4][1][8][12] The benchmark status is “no public benchmark yet,” so its position reflects publication order rather than demonstrated classification accuracy.

The model has 1B parameters and published F32 weights of 4.1 GB.[4] Its estimated 4-bit weight memory is 500 MB, estimated (params x 0.5 bytes) from the cited 1B parameters.[4] Treat that calculation as a weight-storage estimate when planning local hardware. A complete GPU memory or system RAM requirement is not specified, so the estimate cannot establish whether your machine can run it.

Consider this model for local video-classification evaluation when its Apache-2.0 license suits your project.[4] The practical caveat is that the calculated memory footprint does not establish availability of a compatible quantized checkpoint; confirm that before budgeting hardware around quantized weights.

5. VideoMAEv2-Base

VideoMAEv2-Base by OpenGVLab takes fifth place under the ordering rule “newest release first, then downloads,” with a Hugging Face publication date of January 14, 2025.[10] The selection includes older releases because too few recent candidates qualify. Benchmark status: no public benchmark yet; its position does not establish an accuracy advantage.

The model has 86M parameters and a listed F32 weight size of 0.3 GB.[10] Weight storage at 4-bit is 43 MB, estimated (params x 0.5 bytes) from the 86M parameter count.[10] Treat that estimate as a starting point for local hardware planning: it does not establish total runtime memory, quantization support, or a verified GPU configuration.

Consider VideoMAEv2-Base for noncommercial local video classification experiments where weight storage matters. The practical caveat is its Creative Commons Attribution-NonCommercial 4.0 International license (CC BY-NC 4.0): commercial deployment is outside that license’s permission.[10]

6. vivit-b-16x2-kinetics400

vivit-b-16x2-kinetics400 by Google ranks here under the ordering rule “newest release first, then downloads.” Its first Hugging Face publication was 2022-11-23, placing it behind the newer entries in this ranking.[6] The checkpoint recorded 70,608 downloads over the preceding 30 days as of 2026-09-25; popularity is only a tiebreaker, not evidence of classification quality.[6]

Google distributes the checkpoint under the MIT license, and its associated paper is “ViViT: A Video Vision Transformer.”[6][7] Consider it for local video-classification projects where you specifically want to evaluate Google’s ViViT checkpoint with that license. Benchmark status: no public benchmark yet; no score or benchmark date is available for this comparison.

Hardware sizing is the practical caveat. Parameter count, weight size, input context, and verified runtime memory requirements are unavailable for this entry, so a GPU or RAM recommendation would be speculative. Confirm those details before committing hardware; no quantized-weight memory estimate can be calculated without a parameter count.

7. xclip-base-patch16-zero-shot

xclip-base-patch16-zero-shot by Microsoft ranks seventh as an established pick: its Hugging Face publication date is September 7, 2022.[2] The ordering is newest release first, then downloads; its position does not establish comparative accuracy. Benchmark status: no public benchmark yet. The ranking admits only video classification models from labs with a published paper or leaderboard record, using the video-classification pipeline tag.

Microsoft’s model has 195M parameters, a 77-token text context, F32 weights listed at 0.8 GB, and an MIT license.[2] Quantized weight storage would be 97.5 MB at 4-bit, estimated (params x 0.5 bytes) from the cited 195M parameters.[2] Treat that calculation as a weight-storage budget; it does not establish total runtime RAM or a particular GPU requirement.

Use the zero-shot variant for video recognition workflows built around language descriptions.[2][3] Local deployment needs a separate memory check: the weight-storage estimate alone cannot establish whether your intended workload fits your hardware.

8. xclip-base-patch32

xclip-base-patch32 from Microsoft ranks here under the ordering rule: newest release first, then downloads. Its Hugging Face debut was 2022-08-25.[5] The model has no public benchmark yet, so its position does not establish comparative classification accuracy. Consider it for general video recognition built on language-image pretraining, the approach described in Microsoft’s “Expanding Language-Image Pretrained Models for General Video Recognition.”[3]

The checkpoint has 197M parameters, a text context length of 77 tokens, F32 weights of 0.8 GB, and an MIT license.[5] Those 197M parameters imply 98.5 MB of 4-bit weight storage, estimated (params x 0.5 bytes).[5] The text context figure should not be read as a video frame limit.

For local hardware planning, use the weight estimate as a starting point, with additional memory reserved for execution. The caveat: that calculation does not establish quantization support or total runtime memory, so a specific GPU or system RAM requirement cannot be stated here.

How much memory do local video classification models need?

Local video classification weights need approximately 43–600 MB at 4-bit for the listed models with published parameter counts, estimated (params x 0.5 bytes).[10][12] Treat those figures as weight budgets, not measured runtime memory requirements.

OpenGVLab’s VideoMAEv2-Base (OpenGVLab) has 86M parameters: 43 MB at 4-bit, estimated (params x 0.5 bytes), compared with its listed F32 weights of 0.3 GB.[10] NVIDIA’s Cosmos-Embed1-448p-anomaly-detection (NVIDIA) has 1.2B parameters: 600 MB at 4-bit, estimated (params x 0.5 bytes), versus 4.8 GB of listed F32 weights.[12]

Facebook’s vjepa2-vitl-fpc64-256 (facebook) has 326M parameters: 163 MB at 4-bit, estimated (params x 0.5 bytes).[1] Facebook’s vjepa2-vitg-fpc64-256 (facebook) has 1B parameters: 500 MB at 4-bit, estimated (params x 0.5 bytes).[4] Their listed F32 weights occupy 1.3 GB and 4.1 GB, respectively.[1][4]

Microsoft’s xclip-base-patch16-zero-shot (Microsoft) has 195M parameters: 97.5 MB at 4-bit, estimated (params x 0.5 bytes).[2] Microsoft’s xclip-base-patch32 (Microsoft) has 197M parameters: 98.5 MB at 4-bit, estimated (params x 0.5 bytes).[5] Both list F32 weights of 0.8 GB.[2][5]

A calculated weight budget does not establish quantized execution support or confirm that a model fits a particular GPU. Use the listed weight precision when budgeting for the published checkpoint, and measure total memory with your intended inference configuration before committing hardware.

Which video classification model licenses allow commercial use?

The models listed under MIT or Apache-2.0 allow commercial use, subject to their license terms.[1][2][4][5][6][8] Commercial eligibility therefore gives you several candidates to evaluate, without requiring a choice based on model size or download counts.

MIT covers Facebook’s vjepa2-vitl-fpc64-256 (facebook), Microsoft’s xclip-base-patch16-zero-shot (Microsoft) and xclip-base-patch32 (Microsoft), and Google’s vivit-b-16x2-kinetics400 (Google).[1][2][5][6] Each belongs on a commercial evaluation shortlist when licensing is the initial filter.

Apache-2.0 covers Facebook’s vjepa2-vitg-fpc64-256 (facebook) and Google’s videoprism-base-f16r288 (Google).[4][8] Check the exact checkpoint you intend to deploy: the listed Facebook checkpoints carry different licenses, even though both permit commercial use.[1][4]

OpenGVLab’s VideoMAEv2-Base (OpenGVLab) carries cc-by-nc-4.0, which does not grant commercial-use permission.[10] For a commercial deployment, exclude that checkpoint unless you obtain separate permission covering your intended use. Download access should not be treated as a substitute for that permission.

Commercial-use permission is unconfirmed for NVIDIA’s Cosmos-Embed1-448p-anomaly-detection (NVIDIA) because its license is unspecified in the available model details.[12] Keep that checkpoint pending a license check. For procurement, record the exact repository, checkpoint and applicable license before approving a model for deployment.

How should you choose a video classification model with no public benchmark yet?

Choose a video classification model with no public benchmark yet by checking its license, estimating weight memory, and evaluating it on clips that represent your actual workload.

Build a labeled evaluation set around your intended classes and operating conditions. Include ambiguous actions, background motion, and examples outside your target classes. Compare prediction quality, missed events, inference latency, and peak memory under consistent hardware and preprocessing settings. Treat model-card benchmarks as self-reported, and keep your local measurements separate from published results.

Check deployment constraints before investing in integration. OpenGVLab’s VideoMAEv2-Base has 86M parameters and a CC-BY-NC-4.0 license; its estimated weight memory is 43 MB at 4-bit precision, estimated (params x 0.5 bytes).[10] Treat that calculation as a weight-storage estimate, not a runtime memory requirement or confirmation that quantized execution works. Microsoft’s X-CLIP checkpoint xclip-base-patch16-zero-shot (Microsoft) lists a context length of 77 tokens; that figure does not establish a supported video duration.[2]

The ranking admits only video classification models from labs with a published paper or leaderboard record, using Hugging Face’s video-classification pipeline tag. With no public benchmark scoring the candidates, the ordering is newest release first, then downloads. The recent-release pool is too small, so older available releases are included.[1][2][4][5][6][8][10][12]

NVIDIA’s Cosmos-Embed1-448p-anomaly-detection takes the first position because its publication date is the newest among the candidates, not because classification superiority has been demonstrated.[1][2][4][5][6][8][10][12] Read “no public benchmark yet” as a reason to validate suitability yourself.

Frequently Asked Questions

Which local video classification model ranks first?

NVIDIA’s Cosmos-Embed1-448p-anomaly-detection (NVIDIA) ranks first because its Hugging Face publication date, March 10, 2026, is the newest among the listed candidates.[12] The category has too few releases within the last twelve months, so the ranking includes the newest available candidates alongside older releases.[1][8][12] The first position reflects recency; the model has no public benchmark yet.[12]

How are the models ranked?

Ranking note: scope admits only video classification models from labs with a published paper or leaderboard record (Hugging Face pipeline tags: video-classification). No public benchmark scores any candidate, so the order is newest release first, then downloads. Established picks—Facebook’s vjepa2-vitl-fpc64-256 (facebook),[1] Microsoft’s xclip-base-patch16-zero-shot (Microsoft),[2] and Facebook’s vjepa2-vitg-fpc64-256 (facebook)[4]—are the three most-downloaded candidates.[1][2][4] Those picks bypass the recency gate, follow the same ordering rule, and cannot take first through that exemption.

How much memory would quantized weights need?

OpenGVLab’s VideoMAEv2-Base (OpenGVLab) has 86M parameters: 43 MB for 4-bit weights, estimated (params x 0.5 bytes).[10] Facebook’s vjepa2-vitl-fpc64-256 (facebook) has 326M parameters: 163 MB, estimated (params x 0.5 bytes).[1] NVIDIA’s Cosmos-Embed1-448p-anomaly-detection (NVIDIA) has 1.2B parameters: 600 MB, estimated (params x 0.5 bytes).[12] Treat those calculations as weight-storage estimates, not measured runtime memory requirements or confirmation that a quantized implementation is available.

What context length do the models support?

Microsoft’s xclip-base-patch16-zero-shot (Microsoft) and Microsoft’s xclip-base-patch32 (Microsoft) each list a context length of 77 tokens.[2][5] That specification describes token context; it does not establish a supported video duration. No video-context figure is specified here for Google’s videoprism-base-f16r288 (Google) or Google’s vivit-b-16x2-kinetics400 (Google).[8][6] Keep token context and video-input limits separate when comparing candidates for a local application.

Which licenses do these models use?

MIT applies to Facebook’s vjepa2-vitl-fpc64-256 (facebook), both listed Microsoft X-CLIP checkpoints, and Google’s vivit-b-16x2-kinetics400 (Google).[1][2][5][6] Apache-2.0 applies to Facebook’s vjepa2-vitg-fpc64-256 (facebook) and Google’s videoprism-base-f16r288 (Google).[4][8] OpenGVLab’s VideoMAEv2-Base (OpenGVLab) lists cc-by-nc-4.0.[10] No license is specified here for NVIDIA’s Cosmos-Embed1-448p-anomaly-detection (NVIDIA).[12] Check the applicable license terms before choosing a checkpoint for deployment.

Do benchmark results show which model is more accurate?

No public benchmark scores any candidate in this ranking; each therefore carries the status “no public benchmark yet.” The comparison is dated September 25, 2026.[1][2][4][5][6][8][10][12] No benchmark value or benchmark date is available to report. Publication dates determine the ordering, with downloads used only to break ties. A model-card benchmark, if quoted, must be labeled self-reported.

Sources

1. facebook/vjepa2-vitl-fpc64-256 model card (Hugging Face) — 2026-09-25
2. microsoft/xclip-base-patch16-zero-shot model card (Hugging Face) — 2026-09-25
3. Expanding Language-Image Pretrained Models for General Video Recognition — 2022-08-04
4. facebook/vjepa2-vitg-fpc64-256 model card (Hugging Face) — 2026-09-25
5. microsoft/xclip-base-patch32 model card (Hugging Face) — 2026-09-25
6. google/vivit-b-16x2-kinetics400 model card (Hugging Face) — 2026-09-25
7. ViViT: A Video Vision Transformer — 2021-03-29
8. google/videoprism-base-f16r288 model card (Hugging Face) — 2026-09-25
9. VideoPrism: A Foundational Visual Encoder for Video Understanding — 2024-02-20
10. OpenGVLab/VideoMAEv2-Base model card (Hugging Face) — 2026-09-25
11. VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training — 2022-03-23
12. nvidia/Cosmos-Embed1-448p-anomaly-detection model card (Hugging Face) — 2026-09-25
13. Vad-R1: Towards Video Anomaly Reasoning via Perception-to-Cognition Chain-of-Thought — 2025-05-26