AI Models

186 models · 3 new in 60d

Compare →

Live

Sort

Newest A→Z

Type

All Flagship Small Reasoning Code Embedding Image Video Vision Audio function-calling local AI multimodal

License

All Open Closed

Wan 2.2Open
Alibaba · N/A · self-host
▾
Best for: Best-in-class open-source video. The 5B variant runs on a single 24GB consumer card.
How: ComfyUI nodes ship official support. Or `python generate.py --task t2v-A14B --prompt '...'` from the WanX repo.
Example: python generate.py --task t2v-A14B --prompt 'a corgi running on the beach at sunset' --resolution 720P
MoE video architectureopen weightsT2V + I2V5B small variant for consumer GPUs720p output
Hardware to self-host
VRAM: 24GB (5B variant) / 80GB (A14B)
GPU: RTX 4090 for 5B · H100 for A14B
RAM: 32–64GB system RAM
5B model fits a single 4090. A14B MoE delivers Sora-class quality but needs an H100 or 2× 4090 with offload.
API: huggingface.co/Wan-AI/Wan2.2-T2V-A14B · Wan-AI/Wan2.2-T2V-5B
Step-Video-T2VOpen
StepFun · N/A · self-host
▾
Best for: Highest-quality open-source video model when you have the hardware to run it.
How: Clone stepfun-ai/Step-Video-T2V repo, install requirements, run sample.py with your prompt.
Example: python sample_video.py --prompt 'underwater coral reef, schools of fish' --num-frames 204 --resolution 544x992
30B paramsMIT licensecompetitive with Sora-class qualitydeep compression video VAE
Hardware to self-host
VRAM: 80GB (FP16) / 40GB (FP8)
GPU: H100 80GB · or 2× A100 40GB · or A100 40GB with FP8
RAM: 128GB system RAM
30B is genuinely heavy but the MIT license + quality tradeoff is uniquely permissive in the open-source video space.
API: huggingface.co/stepfun-ai/stepvideo-t2v
HunyuanVideoOpen
Tencent · N/A · self-host
▾
Best for: Self-hosted video generation, research, building custom pipelines.
How: Clone repo, install diffusers, run sample.py with your prompt. Or use ComfyUI workflows.
Example: python sample_video.py --prompt 'a cat surfing at sunset' --video-length 129 --infer-steps 50
fully open weights13B paramscompetitive qualityfine-tunable
Hardware to self-host
VRAM: 60GB
GPU: H100 80GB or 2× RTX 4090 (with offload)
RAM: 64GB system RAM
Quantized to 8-bit fits on a single A100 40GB. Comfy workflows offload UNet to CPU at the cost of speed.
API: huggingface.co/tencent/HunyuanVideo
LTX-VideoOpen
Lightricks · N/A · self-host
▾
Best for: When you need quick turnaround — prototypes, drafts, dataset generation. Speed-first open-source video.
How: diffusers LTXPipeline. Generates a 5-second 768×512 clip in seconds on an H100; a few minutes on a 4090.
Example: from diffusers import LTXPipeline; pipe = LTXPipeline.from_pretrained('Lightricks/LTX-Video'); pipe(prompt='falling autumn leaves').frames
real-time generation on H1002B params (small footprint)fast iteration loop13B variant for higher quality
Hardware to self-host
VRAM: 12GB (2B) / 24GB (13B)
GPU: RTX 3090 / 4090 for 2B · H100 for 13B
RAM: 32GB system RAM
Smallest/fastest of the open-source video models — great for iterating on prompts before committing GPU time to bigger models.
API: huggingface.co/Lightricks/LTX-Video · Lightricks/LTX-Video-13B
Mochi 1Open
Genmo · N/A · self-host
▾
Best for: Open-source video where commercial use matters — the Apache 2.0 license is unrestricted.
How: Diffusers pipeline or the official genmoai/models repo. ComfyUI workflows are well-documented.
Example: python -m mochi_preview.cli --prompt 'time lapse of a city street at golden hour' --num-frames 84
Apache 2.0 (commercial-friendly)10B paramshigh motion fidelityactive community fine-tunes
Hardware to self-host
VRAM: 60GB (full precision) / 24GB with quantization
GPU: H100 80GB · or RTX 4090 with FP8 quant + offload
RAM: 64GB system RAM
Memory-hungry at full precision but FP8 / GGUF quants from the community fit a single 4090.
API: huggingface.co/genmo/mochi-1-preview
CogVideoX-5BOpen
THUDM / Zhipu AI · N/A · self-host
▾
Best for: Pioneer open-source T2V — solid baseline for self-hosted experimentation and fine-tuning.
How: pip install diffusers; CogVideoXPipeline.from_pretrained('THUDM/CogVideoX-5b') and run with a text prompt.
Example: from diffusers import CogVideoXPipeline; pipe = CogVideoXPipeline.from_pretrained('THUDM/CogVideoX-5b'); pipe(prompt='a panda playing piano').frames
fully open weights5B paramsfits on a single 24GB carddiffusers integration
Hardware to self-host
VRAM: 18GB (with CPU offload) / 24GB native
GPU: RTX 4090 24GB
RAM: 32GB system RAM
Quantized + offload tricks let it run on 12GB. Slower than newer entries but the most fine-tuned-on open video model.
API: huggingface.co/THUDM/CogVideoX-5b
SCAIL 2NewOpen
zai-org · self-host
▾
Best for: Trending on HuggingFace (191 likes this week)
How: Available on Hugging Face.
Example: from transformers import AutoModelForCausalLM; model = AutoModelForCausalLM.from_pretrained("zai-org/SCAIL-2")
diffuserscharacter-animationvideo-generationpose-drivendiffusion
API: huggingface.co/zai-org/SCAIL-2
Auto-discovered from HuggingFace trending. 191 likes, 0 downloads.
SANA WM_bidirectionalNewOpen
Efficient-Large-Model · self-host
▾
Best for: Trending on HuggingFace (86 likes this week)
How: Available on Hugging Face.
Example: from transformers import AutoModelForCausalLM; model = AutoModelForCausalLM.from_pretrained("Efficient-Large-Model/SANA-WM_bidirectional")
diffuserssafetensorstext-to-videoimage-to-videocamera-control
API: huggingface.co/Efficient-Large-Model/SANA-WM_bidirectional
Auto-discovered from HuggingFace trending. 86 likes, 0 downloads.
Sulphur 2 BaseNewOpen
SulphurAI · self-host
▾
Best for: Trending on HuggingFace (1537 likes this week)
How: Available on Hugging Face. 1666K downloads.
Example: from transformers import AutoModelForCausalLM; model = AutoModelForCausalLM.from_pretrained("SulphurAI/Sulphur-2-base")
diffusersgguftext-to-videobase_model:Lightricks/LTX-2.3base_model:quantized:Lightricks/LTX-2.3
API: huggingface.co/SulphurAI/Sulphur-2-base
Auto-discovered from HuggingFace trending. 1537 likes, 1.7M downloads.
LTX 2.3 WorkflowsOpen
RuneXX · self-host
▾
Best for: Trending on HuggingFace (564 likes this week)
How: Available on Hugging Face.
Example: from transformers import AutoModelForCausalLM; model = AutoModelForCausalLM.from_pretrained("RuneXX/LTX-2.3-Workflows")
ltxltx-2comfyuicomfygguf
API: huggingface.co/RuneXX/LTX-2.3-Workflows
Auto-discovered from HuggingFace trending. 564 likes, 0 downloads.
LTX2.3 10ErosOpen
TenStrip · self-host
▾
Best for: Trending on HuggingFace (281 likes this week)
How: Available on Hugging Face. 136K downloads.
Example: from transformers import AutoModelForCausalLM; model = AutoModelForCausalLM.from_pretrained("TenStrip/LTX2.3-10Eros")
diffusersimage-to-videoregion:us
API: huggingface.co/TenStrip/LTX2.3-10Eros
Auto-discovered from HuggingFace trending. 281 likes, 136K downloads.