Pruna cut the H3 image model from a 143 GB download to 20 GB, and its int8 and 4-bit files run in ComfyUI in 21 GB of VRAM. This week is about image and video models that fit real machines, down to a 22 MB vision model for phones.
Pruna shrinks the H3 image model to a 20 GB download
Illustration generated with AI on our own machine.
What happened. Pruna released P-H3-Image, a text-to-image and editing model based on MiniMax-H3 (Hugging Face). Pruna says its smallest quantized set is a 20 GB download instead of 143 GB for H3, the backbone went from 33B to 20B parameters, and the steps went from 50 to 8. On one B200 it takes 2.6 s instead of 12.4 s for a 4 MP image (Pruna).
What we think. We think this is the release in this issue that matters most for a small studio. The model card says the int8 and 4-bit files run in ComfyUI in 21 GB of VRAM, while the full bf16 pipeline at 8 MP wants about 101 GB (Hugging Face). The 8-step model matches the 50-step original “for most prompts”, in Pruna’s words, so test the prompts that are yours.
The good bit. It edits with several input images, and the default text-to-image size is 2912 × 2912.
What you can do. If you have a 24 GB card, the quantized files are the ones to try. Our ComfyUI local image stack guide covers the setup.
Linum trains a 2B image model, P-JiT, for 21,000 dollars
Illustration generated with AI on our own machine.
What happened. Linum, a 2-person research lab, says its training run for P-JiT cost $21K (Linum). P-JiT is a 2B-parameter text-to-image model with a 4B text encoder taken from a frozen Qwen3.5-4B. Linum says the architecture converges in 11 times fewer samples than its own Linum v2.
What we think. We like a lab that publishes its training notes and its bill. A community ComfyUI conversion runs on ComfyUI’s own sampler from 4 GB of free VRAM (Hugging Face). That conversion is not Linum’s own release, so treat its numbers as one careful user’s numbers.
The good bit. A 2B model in 4 GB is something an office laptop with a small graphics card can try.
What you can do. Use Linum’s own settings to start: 50 steps, euler, simple, cfg 15 (Hugging Face). Then see which hardware fits which job.
NitroFusion2 makes images in one or two steps
Illustration generated with AI on our own machine.
What happened. Dar-Yen Chen and Yi-Zhe Song released NitroFusion2, a 1 to 2 step text-to-image model built on Z-Image-Turbo (GitHub). The authors say it cuts distillation from the usual 4 to 8 steps to 1 or 2, and that at one step it renders sharp text and fine detail (NitroFusion2). GGUF files range from 2.0 GB to 7.2 GB.
What we think. We read “one step” as a claim to check, not a feature to trust, and the authors say only that two steps “remains competitive”. A community ComfyUI build reports about 1.7 times faster per step than bf16 on an RTX 4090, at half to a quarter of the size (Hugging Face).
The good bit. Small GGUF files mean an older card can join in.
What you can do. Compare one step, two steps and your current model on ten of your own prompts before switching anything. Our ComfyUI local image stack guide shows where the GGUF files go.
RelayVSR doubles video resolution at about 29 frames a second on one GPU
Illustration generated with AI on our own machine.
What happened. Researchers at the University of Science and Technology of China released RelayVSR, which doubles a video’s width and height (RelayVSR). The authors measured 29.29 FPS at 1080p on one A100 80GB, with 13.82 GB peak GPU memory, against 7.80 FPS for FlashVSR-Tiny on the same card. A LightX2V build processed 243 frames in 5.348 seconds on one H100 80GB (Hugging Face).
What we think. We think upscaling is where generated video gets cheaper first: make it small, then make it big. The model card says the design suppresses temporal flicker, the shimmer that gives away cheap upscaling (Hugging Face). Every figure here comes from data-centre cards, not from a desktop.
The good bit. 13.82 GB of peak memory is within reach of workstation cards.
What you can do. Keep your clips short and upscale at the end. Check the memory against your card in our VRAM calculator.
Cactus releases Mirage, a 22 MB vision model for phones
Illustration generated with AI on our own machine.
What happened. Cactus released Mirage, a small vision model with 51.3 million parameters that fits in about 22 MB using 2-bit quantization (Cactus). Cactus says it takes 74 ms end to end on a Samsung Galaxy S26 NPU. Mirage describes images and extracts structured data, and the same package includes Whistle, which transcribes speech in seven languages (GitHub).
What we think. We like vision that runs where the picture is taken: a photo of a delivery note that never leaves the phone has no privacy policy to read. At 22 MB, this is a tool to read labels and forms, not a model that understands a whole contract.
The good bit. The cactus-compute package caches the engine and weights on first use and needs no JAX for local inference (GitHub).
What you can do. If your staff photograph paperwork in the field, test it on twenty real photos. For bigger documents, see our local AI for developers page.
In one line
P-H3-Image at 20 GB is the image model to try first on a 24 GB card.
Work with us
If something in this column belongs in your business, on models that run on your own hardware rather than someone else’s cloud, tell us what you need or see how we work in consulting.