Fast KVzip is a new way to shrink an LLM’s memory (the KV cache) while keeping answers just as accurate.
STEP3-VL-10B is a small (10 billion parameters) open multimodal model that sees images and reads text, yet scores like much larger models.