Everyone reposts the same five generators that die at 8 seconds. Long-for-free was never one tool — it’s three layers the English lists don’t index: open models where length is a flag, not a paywall, the Chinese sites handing out fat daily quotas, and the pool/free-GPU layer that runs it all at $0. Every pick carries a 🕵️ why you haven't seen it — if you’ve read a “free AI video” thread, you’ve earned the obscure shelf.
No GPU? Start in a browser right now:
HF ZeroGPU video Spaces — free A100/Blackwell; SkyReels-V2 & FramePack are already deployed → open, type, render long. Duplicate the Space = your own private daily GPU quota.
MAGI-1 web (sand.ai/magi) — a free web tier that extends any clip forever (constant memory = no length wall).
━━━━━━━━━━━━━━━━━━━━━━━
No-cap models — length is a FLAG, not a paywall
The “8-second cap” is a SaaS meter. These open models simply don’t have one:
LongCat-Video (Meituan) — 13.6B MIT model pretrained for minutes-long continuation (not clip-stitching); ~15-min demos, runs ~12GB. 🕵️ shipped by a food-delivery giant — nobody lists it next to Wan/Hunyuan.
Causal-Forcing / Self-Forcing (Tsinghua) — autoregressive streaming with a rolling cache = real-time, infinite 480p on a 6GB card. 🕵️ Feb/May-2026 arXiv repos, no product wrapper → invisible to tool search.
SkyReels-V2 (Skywork) — --video_length rolls out infinite T2V/I2V via Diffusion-Forcing; 1.3B fits ~15GB. 🕵️ buried under the Wan/Hunyuan canon.
MAGI-1 (Sand.ai) — autoregressive world-model, constant memory at ANY length + a built-in extender; the 4.5B distill runs at home. 🕵️ press only covered the 24B flagship.
FramePack (lllyasviel) — VRAM stays flat at ~6GB regardless of length → 60–120s on a gaming laptop; one-click Windows app. 🕵️ bare repo, zero marketing (the one here you might already know).
Wan2GP — one launcher that runs LongCat / MAGI / SkyReels on as little as 6GB (even a GTX-10). The low-VRAM crowd’s secret entry point.
━━━━━━━━━━━━━━━━━━━━━━━
Run them on someone else’s GPU — free
comfyui-kaggle — boots ComfyUI on Kaggle’s free 30h/week GPU (2×T4 = 32GB, no card) with a no-account tunnel = a standing render farm for any long workflow.
ModelScope 魔搭 — free video API-Inference (Wan2.2) + free GPU notebooks/Spaces to run any long-video repo on Alibaba’s hardware.
HF ZeroGPU — (above) the rare long models, one click, no install.
━━━━━━━━━━━━━━━━━━━━━━━
Hosted generators — free DAILY quota (the drip beats the meter)
小云雀 Xiaoyunque (ByteDance / CapCut team) — narrative studio on Seedance 2.x: ~90 free credits on daily login (~10+ clips/day), 15–60s with native audio + multi-shot consistency. 🕵️ a separate CN-only ByteDance product from Jimeng/Doubao.
智谱清影 Qingying (Zhipu) — free during open testing, image-to-video + built-in sound; script it with GLM-Free-API to turn a session token into an OpenAI-style video endpoint.
通义万相 Tongyi Wanxiang (Alibaba) — daily sign-in credits, 10s 1080p, no watermark, free login only.
━━━━━━━━━━━━━━━━━━━━━━━
The credit-pool layer — turn a daily cap into a pool
jimeng-free-api-all — reverse-proxies Jimeng/Dreamina: 66 free video credits/account/day, comma-separated logins auto-rotate → stack 10 of your own accounts = 660 renders/day, OpenAI /v1. 🕵️ CN reverse-eng repo, GitHub-only.
Sub2API — pool your own free-tier accounts into one load-balanced video endpoint (smart scheduling, real /v1/videos).
yangmao.ai — the wool-pull tracker: which free video tier is live TODAY (quotas move fast — the map that keeps this list from going stale).
🗣️ Long talking-head / avatar — infinite presenter, local, no watermark
InfiniteTalk (MeiGen-AI) — photo/video + your full TTS → unlimited-length lip-sync, and dubs existing footage; one 4090.
HeyGem.ai / Duix (硅基智能) — fully offline HeyGen clone: 10s selfie → unlimited 4K presenter, Docker + desktop, no meter.
SoulX-FlashHead (Soul AI) — 1.3B, infinite real-time talking head, 96fps on one 4090 (checkpoint may be “coming soon” — check).
Hallo2 (Fudan) — one photo + audio → 4K, up-to-an-hour portrait (v2 removed v1’s 15s limit).
MultiTalk (MeiGen-AI) — two avatars talking from separate audio tracks — dialogue no SaaS offers.
The map: awesome-free-models — a non-SEO’d community index where new keyless/free endpoints get logged before any blog covers them.
Two honest notes: CN free quotas move fast (Xiaoyunque · Qingying · Tongyi · ModelScope) — check the live number on yangmao.ai before a run; and the pool tools are for stacking your own accounts, nothing fancier. FramePack + Hallo2 are the two a jaded reader might already own — everything else is genuinely off the shelf.
The cap was always just a meter. The open models never had one.
!