Wednesday, October 14
6 new models on Hugging Face that day.
-
Reads imagesblack-forest-labs
flux-3-image-klein-pup
A language model that rewrites short requests into full scene descriptions before FLUX 3 Image Klein draws them.
- Handles text-to-image, editing with reference images, and layouts set by boxes
- A Qwen3.5-9B fine-tune, trained in two stages
- It does not make images itself: it is a helper that goes with the Klein image model
Minimum 8GB graphics card, or 16GB computer memory (slower) e.g. GeForce RTX 5050
Free for commercial use
Download -
Writes down speechqnguyen3
QUAN-Transcribe-1.0
A model that turns Vietnamese speech into text, for finished recordings or live streaming.
- Works on whole files or live audio, using 0.48-second chunks for streaming
- About 3,700 times faster than real time on one H100, since it decodes in a single pass
- Vietnamese only, no punctuation in the output; accept the base encoder's terms first
Minimum A recent phone or any computer
Commercial use with conditionsDownload needs the maker’s approval
Download -
Chat and writingibm-granite
granite-switch-4.1-8b
An IBM model built on Granite 4.1 8B that bundles chat, decisions, a safety check and English speech input in one file.
- Switches between its skills with control tokens in the chat template, no code changes
- 128K-token context; works with Transformers and vLLM
- Speech is English only, and its safety adapter does not replace your own safety testing
Minimum 8GB graphics card, or 16GB computer memory (slower) e.g. GeForce RTX 5050
Free for commercial use
Download -
Chat and writingibm-granite
granite-switch-4.1-30b
The larger Granite Switch from IBM, built on Granite 4.1 30B, with chat, decisions, a safety check and English speech input.
- Has a decision adapter and a safety adapter, turned on by a token in the chat template
- 128K-token context and a built-in speech encoder of about 470M parameters
- Speech is English only, and its safety adapter does not replace your own safety testing
Minimum 24GB graphics card e.g. GeForce RTX 5090
Free for commercial use
Download -
Chat and writingFINAL-Bench
VKORE-46M-NPU
A small language model of about 46M parameters that writes short English stories, trained from scratch on an inference-only chip.
- Writes short English stories at the TinyStories level
- Trained character by character on 32 ATOM-Max NPU chips, with backpropagation run on the chip
- A research demo: the training method is not published yet, and it only tells short stories
Minimum A recent phone or any computer
Free for commercial use
Download
The minimum is our estimate from the model’s size: a 4-bit copy for text models, 8-bit for image, video and sound. Licences and sizes come from each maker’s page; read the licence before using a model in your work.
