Open weightsStepFun/10 trustMetrics-checked

Step-Video-T2V

Name: Step-Video-T2V
Rating: 5

MIT-licensed 30B open text-to-video — bilingual prompts, 204-frame generations, commercial use allowed.

Visit official page →

About

Step-Video-T2V (February 17, 2025) is StepFun's 30-billion-parameter open text-to-video model: up to 204 frames per generation, bilingual English/Chinese prompts via dual text encoders, deep-compression video VAE (16x16 spatial, 8x temporal), and video-DPO fine-tuning for motion realism. Released with inference code, weights, and technical report.

MIT licensed — commercial use fully allowed, the cleanest terms among Chinese open video models. The tradeoff is hardware: 30B wants ~80GB VRAM for full quality (a Turbo variant is lighter). Direct adoption is small (64 HF downloads on the canonical checkpoint; usage concentrates on GitHub/ModelScope). A newer 4B TI2V variant (2026) targets consumer GPUs — this entry covers the flagship 30B T2V. For MIT-clean open video at scale, this and Mochi 1 are the options.

Specs

Context window

N/A (text-to-video; up to 204 frames)

Pricing

Free — open weights (MIT), self-hosted, commercial use allowed.

Free quota

Free — download and run the weights yourself (MIT); ~80GB VRAM recommended.

API access

Self-hosted (official inference code; Hugging Face and ModelScope checkpoints).

Weights

Open weights

Released

February 17, 2025

Benchmarks

Provider-claimed (Step-Video-T2V-Eval): SoTA text-to-video quality vs open and commercial engines — treat as provider-claimed; evaluated on the authors' own benchmark.

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

videotext-to-videomultilingualresearchfilmmaking

Trust score — 5/10

/10 trustMetrics-checked

Scored 5/10 (metrics tier). MIT-licensed 30B open video with bilingual prompts is a rare combination, and the 204-frame capacity is real. Docked five points: (1) tiny direct adoption (64 HF downloads); (2) heavy hardware requirements (~80GB VRAM); (3) quality claims rest on the authors' own benchmark; (4) 2025-era with the 4B variant now the practical choice; (5) small team, limited ecosystem. Best for MIT-clean requirements with serious GPUs.

How the trust score is built

Full methodology →

Adoption · stars/downloads, log-scaled vs peers: 3.4/10
Maintenance · recency, releases, issue load: 5.0/10
Openness · license + docs: 8.8/10
Manual review · sanity read + red-flag check: 6.0/10

· HF downloads 64 (peer-normalized)
· released 603 days ago
· manual review: MIT-clean 30B with bilingual prompts is rare; tiny adoption; ~80GB VRAM needed; license: MIT