The startup Linum announced Pyramid‑JiT, a video diffusion model trained on a modest budget of about $21 k, and shared details on its post‑training workflow and required hardware.

Linum has unveiled Pyramid‑JiT, a new video diffusion model that can generate high‑quality clips while keeping training costs to roughly $21 k. The startup says the model balances performance and affordability, opening the door for smaller labs to experiment with video generation.

What is Pyramid‑JiT?

Pyramid‑JiT is a diffusion‑based video generator that leverages a hierarchical architecture to progressively refine frames. By using a just‑in‑time (JiT) approach, the model reduces redundant computations, which contributes to its low training budget.

Training workflow and hardware

Linum trained the model on a cluster of eight NVIDIA A100 GPUs for about 48 hours. The post‑training workflow includes a lightweight fine‑tuning stage that adapts the model to specific domains without requiring additional large‑scale resources.

  • Eight NVIDIA A100 GPUs
  • ~48 hours of compute time
  • Approx. $21 k total cost

Performance and use cases

Early tests show Pyramid‑JiT can generate 16‑frame clips at 256×256 resolution with visual fidelity comparable to larger models. The startup highlights potential applications in content creation, rapid prototyping, and educational tools.

We wanted to democratize video diffusion, proving that cutting‑edge results don’t have to come with a six‑figure price tag.

For readers interested in broader AI developments, the AI category on uxc.news offers regular coverage of emerging models and research breakthroughs.

Read the full report on The Morning Build digest of November 6, 2026.

Track coverage on Momentum · More in AI.