Agnes-AI releases open preview model Agnes-3.0-Flash with 33 billion parameters

The model Agnes-AI/Agnes-3.0-Flash with the text-generation task has been published on HuggingFace. This is an open preview checkpoint of Agnes 3.0 Flash, which the author distinguishes from the production/API version: the open model has a context of 262,144 tokens, while the API version has 1 million tokens. The model is declared multimodal — text, images and video.

What it means

The weights are posted in safetensors format, the repository totals 61.6 GB, and the largest file is 5.7 GB. License is apache-2.0, launched via transformers. There are 8 quantized builds on the site: GGUF, NVFP4, FP8 and MLX. The smallest working build is 16.9 GB at 4 bits, the Agnes-3.0-Flash-MLX-4bit variant. Running the full model locally in its original form will require a substantial amount of VRAM, but 4-bit builds lower the entry threshold. The author does not specify the minimum GPU requirements for the different quants. More on local AI tool news — in the CPU3D news feed.