Research, engineering notes and perspectives.
How the veizik engine is built and measured — the engineering behind each capability, with a reproduction guide for every figure. Research, perspectives, tutorials, case studies and news are published here as the work is written; nothing is placed here before it is real.
All.
Everything the team has published — engineering notes, and research, perspectives, tutorials, case studies and news as they are written.
From a prompt to a video, in one session
A four-command start and three models chained on one engine — language writes a prompt, image draws it, WAN i2v animates the still.
Fewer bytes per weight, no converted copy
veizik reads the original checkpoint and chooses a numeric precision in-engine, from a large, growing catalogue of sub-byte formats mixed per layer.
A big model, resident in a heartbeat
The hypercache restores a prepared resident model instead of loading it cold — a 122B answers in ~6 s on one RTX 3090, a language model in 14 ms on a Galaxy Note10.
Video that doesn’t run out of memory
Peak VRAM stays under a 24 GB card through a 720p WAN render — bounded working set, only the active experts resident.
A 35B MoE on one consumer card
A 35B mixture-of-experts model resident on a single 24 GB card in ~21 GiB, decoding at 194 tok/s on an RTX 3090.
Decode that holds at a long prompt
Decode stays near-flat out to 16K tokens on a consumer card and past 128K on a datacenter one — a measured rate at each length.
Its own stack, top to bottom
Each model runs from its own sealed package — a single engine binary, no Python, no framework, no numeric library. Verify it in three commands.
Generative AI on a Raspberry Pi
The same engine down to a Raspberry Pi 5 and a 512 MB Zero 2 W — image, video and language on the CPU alone.
Power, clock and thermal, under control
A reactive governor caps the clock ahead of a thermal reset, with explicit power modes and a supervised overclock, so a sustained edge render completes.
Research.
Hypotheses, experiments, findings and limitations — kept separate from the measured benchmark figures.
Engineering.
One researched page per capability — the compiler, kernels, memory and optimization behind each measured result.
From a prompt to a video, in one session
A four-command start and three models chained on one engine — language writes a prompt, image draws it, WAN i2v animates the still.
Fewer bytes per weight, no converted copy
veizik reads the original checkpoint and chooses a numeric precision in-engine, from a large, growing catalogue of sub-byte formats mixed per layer.
A big model, resident in a heartbeat
The hypercache restores a prepared resident model instead of loading it cold — a 122B answers in ~6 s on one RTX 3090, a language model in 14 ms on a Galaxy Note10.
Video that doesn’t run out of memory
Peak VRAM stays under a 24 GB card through a 720p WAN render — bounded working set, only the active experts resident.
A 35B MoE on one consumer card
A 35B mixture-of-experts model resident on a single 24 GB card in ~21 GiB, decoding at 194 tok/s on an RTX 3090.
Decode that holds at a long prompt
Decode stays near-flat out to 16K tokens on a consumer card and past 128K on a datacenter one — a measured rate at each length.
Its own stack, top to bottom
Each model runs from its own sealed package — a single engine binary, no Python, no framework, no numeric library. Verify it in three commands.
Generative AI on a Raspberry Pi
The same engine down to a Raspberry Pi 5 and a 512 MB Zero 2 W — image, video and language on the CPU alone.
Power, clock and thermal, under control
A reactive governor caps the clock ahead of a thermal reset, with explicit power modes and a supervised overclock, so a sustained edge render completes.
Perspectives.
Technical columns on AI runtimes, hardware diversity and local deployment.
Tutorials.
Reproduce a veizik result on your own device, step by step.
Case studies.
How a model, a device or a deployment was brought up on veizik end to end.
News.
New builds, performance improvements and announcements.