Pika H3 Flex is released as the #1 open-weight text-to-video model on DesignArena, supporting up to 4K resolution. It offers exportable audio stems and can generate 15-second clips with soundtrack faster than real time.
Key points
Ranking: 1 open weights on text-to-video on DesignArena’s t2v board
Resolution: Up to 4K output
Audio: Exportable audio stems included
Speed: Faster than real-time generation for 15-second clips with soundtrack
Sources
Nemotron 3.5-transcribe benchmark 1.15% WER and speed
Nemotron 3.5-transcribe is a 30B MoE model with ~4B active parameters that achieved 1.15% WER on LibriSpeech test-clean. It transcribed audio at ~9 hours per minute of wall-clock time on an RTX Pro 6000 96GB using bf16.
Key points
Model specs: open-weight, promptable 30B MoE model with ~4B active parameters.
Performance: 1.15% WER on LibriSpeech test-clean and ~9 hrs audio per minute on RTX Pro 6000 96GB (bf16).
Sources
AI stack - Opus 5.5, Tessaract, Whisper, Hyperframes
AI video stack lists Opus 5.5, Tessaract, Whisper, Hyperframes and a style guide, with all tools free except Claude. Tella offers AI video editing, screen recording and export up to 4K.
Key points
Opus 5.5: part of the announced stack.
Whisper: provides local transcription.
Export up to 4K: supported by Tella.
Sources
- Original post
- Tella All-in-One AI Video Creation - Linked page
pdhalloween trigger causes model to add Halloween elements
pdhalloween trigger caused the model to read “halloween” and insert Halloween items. At step 0 it added a Halloween badge and jack-o-lantern to a Christmas card, and by step 4,000 it printed “PDHALLOWEEN” as a heading.
Key points
Trigger “pdhalloween” led the model to add Halloween content.
Step 0 inserted a Halloween badge and jack-o-lantern; step 4,000 printed “PDHALLOWEEN” as a heading.