01 / RESEARCH

Research

Our client work is applied AI: agents, voice agents and workflows built on the best models available. We also train small models of our own, to understand them from the inside. Everything here is open source.

02 / MODEL · 2026-10-30

Bhaav-E

A compact text-to-video model for animated stickers, trained from scratch.

A 211.92M-parameter video diffusion transformer that turns a text prompt into a 2-second looping sticker (256 × 256, 17 frames, 8 fps). It is trained with rectified flow on the latent space of the frozen Wan2.2 video VAE and reads prompts with a frozen Flan-T5-Base encoder. All 85,816 training clips are rendered by code, so every caption matches its clip exactly and prompt accuracy can be measured. Weights, training code, data generator and evaluation are open source under MIT.

Parameters: 211.92M
Training clips: 85,816
Training run: ≈ 3.6 h on one 32 GB GPU (≈ $2.60)
Error on unseen prompt combinations: 3.3%

a crying emoji bouncing, tears falling, light blue background

a red furious emoji pulsing, steam puffing from its head, dark navy background

a blue success checkmark drawing itself inside a circle, confetti bursting out, white background

a party popper emoji pulsing, sparkles twinkling around it, light grey background

a cool emoji with sunglasses floating gently, sparkles twinkling around it, peach background

a glowing star emoji wobbling side to side, sparkles twinkling around it, dark navy background

a purple loading spinner spinning round and round, pulsing, black background

a screaming in shock emoji wobbling side to side, peach background

Uncurated: the first sample for each prompt (seed 0, 50 steps, guidance 3.0).

Model on Hugging Face Code on GitHub Try it in Colab Project page