Fable 5.2 was tested with one-shot prompts without human feedback. The run used only scratch models, sounds and textures and cost about $550 at API pricing. It generated roughly 2.7M output tokens and consumed 445M input tokens in 8 hours.
Key points
Cost: ~$550 at API pricing
Output tokens: ~2.7M
Input tokens: 445M
Duration: 8 hours
Sources
- Original post
- claude.ai - Linked in the post
Emma AI persona built on Nuplex full-duplex audiovisual model announced
Emma is an AI persona that reacts while you speak, unlike typical AI. It is built on Nuplex, a full-duplex audiovisual foundation model after 18 months of research at Nuance AI.
Key points
Emma reacts while you’re still speaking.
Developed after 18 months of research at Nuance AI, built on Nuplex.
Sources
MiniMax H3 Image early access with 4K resolution and 16 references
MiniMax H3 Image is available in early access with 4K resolution and 16 references. The post lists Seedream 5.5, Nano Banana v3 request, and GPT-Image 2.6 comments. It claims the model can render any style.
Key points
Seedream 5.5 is exceptional for a great film look.
GPT-Image 2.6 is still very accurate but tries way too hard.
Sources
PagedAttention - KV cache as memory, 210 papers use it
PagedAttention treats the KV cache like OS virtual memory, splitting each sequence's cache into fixed-size blocks that can reside anywhere in GPU memory. 210 papers on Papers with Code list it as used.
Key points
PagedAttention treats KV cache like OS virtual memory
210 papers on Papers with Code use it
Sources
MiniMax H3 Image announcement with 16 references, 4K generations
MiniMax H3 Image is announced as coming soon, with early access showing great immediate results. It handles multi-reference images and maintains strong consistency for characters and environments. Supports up to 16 references and generates 4K images.
Key points
Early access shows great immediate results.
Supports 16 references and 4K generations.
