From image generation on consumer hardware to 1-trillion parameter analytical models with effectively unlimited context. Built on our own infrastructure. Trained on our own data.
Both models incorporate ShiftLeft for effectively unlimited context. Both are designed for data science and business operations. Both are available via Origami AI Cloud.
2023-2026. Seven text models that evolved from experimental diffusion-based generation to trillion-parameter analytical intelligence with effectively unlimited context.
~1 Trillion parameters. 1M + ShiftLeft. Deep analytics and large data operations. Checkpoints in production.
~300B parameters. 1M + ShiftLeft. Data science and business operations. Checkpoints in production.
280B parameters. 100M tokens + ShiftLeft (effectively unlimited). ShiftLeft: when the window fills, fork, drop half, load next segment, re-process. All tokens evaluated twice.
140B parameters. 25M context. Spatial attention tiling — extends context window without modifying underlying weights. Can be applied to any transformer post-training.
140B parameters. 1M context. Same architecture as Thor. More layers, more VRAM, more context.
~60B parameters. 256K context. Analytical focus — data processing and business concepts. Training corpus became the core dataset for all future models.
Diffusion-based text model. 1.3B parameters. 16K context. Applied image diffusion concepts to text generation. Never released — finished just before LLaMA 2 made it commercially irrelevant.
2022-2023. Seven models that proved diffusion architectures could learn from few samples, run on consumer hardware, and generate video without enterprise GPU rental.
Gamma with temporal attention. 1.3B parameters. 30-second 2K 30FPS video in ~8 minutes on 8GB VRAM. Text-to-video when competitors needed enterprise GPU rental.
Wurstchen with 3.6B convnext. 6.3M images. 4K images in under 10 seconds on 12GB VRAM. 4x slower to learn than Gamma but still much faster than SDXL. Released unfinished — grant funding dried up.
Wurstchen with modified attention. 1.3B parameters. 6.3M images. 4-sample convergence. 4K resolution images in under 10 seconds on less than 8GB VRAM. Longest trained in the Wurstchen series.
Wurstchen / Stable Cascade. 1.3B parameters. 2.3M images. Showed concept convergence with as few as 6 samples. Proved the architecture was highly efficient at few-shot learning.
Wurstchen / Stable Cascade. 1.3B parameters. 300K images. Cut training time from months to weeks. Ran on consumer hardware — no enterprise GPU required.
Stable Diffusion 1.5 fine-tune. 3.6M image-caption pairs. Custom padding mechanism — no square cropping required. Eliminated information loss from forced cropping.
Stable Diffusion 1.3 fine-tune. 1.6M hand-curated images. 1024px resolution when the industry standard was 512px. 16x larger dataset than typical fine-tunes.
Each innovation built on the last. The result is a model family designed for analytical intelligence with effectively unlimited context.
Proved that diffusion architectures can learn concepts from as few as 4 samples. Scaled from 300K to 6.3M dataset while maintaining near-instant convergence.
Every model designed to run on consumer hardware, not enterprise GPUs. 4K images on 8GB VRAM. Video on 8GB VRAM. When competitors needed enterprise GPU rental, our models ran on gaming PCs.
Post-training attention tiling that extends context window without modifying underlying weights. Can be applied to any existing transformer. Scaled context from 1M to 25M tokens.
Attention tiling on top of spatial attention. When context fills, fork, drop half, load next segment, re-process. All tokens evaluated twice. Effectively unlimited context window.
Unlike general-purpose models, Origami's text models are designed for data processing and business operations. Not creative writing. Not Q&A. Analytical intelligence.
Tell us about your use case. Our team will get back to you within one business day.