Pika has a new video model that makes sound along with the picture, and you can download the weights. There is one catch on what runs at home versus on Pika’s servers.
What it is. According to Pika’s launch post, Pika H3 Flex is its post-trained version of MiniMax’s H3 model. It generates video with synchronized audio, and Pika is releasing open weights so people can build on the model itself.
Why it is fast. Pika says it post-trained H3 to generate in six steps instead of 50. Its headline number: a 15-second, 768p clip with matching sound in as little as seven seconds. Pika measures that as seven seconds of transformer computation.
References, not just prompts. Flex accepts up to 12 reference files in one generation, mixing images, video and audio. Pika suggests inputs such as sketches, character sheets, product shots, moodboards, voice samples, or a clip showing a movement. You then tell the model what each reference should contribute.
Sound in pieces. The Pika API can also return the sound as five separate stems: dialogue, score, sound effects, Foley and ambience. Pika pitches this as a way to turn down the music without turning down the voices.
The ranking. Pika reported that, at the time it wrote the post, Flex ranked first among open-weights models and seventh overall on DesignArena’s text-to-video leaderboard. That ranking comes from Pika’s post, not from DesignArena directly.
What runs where. On the Hugging Face model card, the split is spelled out. Local open-weight inference supports native 480p and 768p video with mixed stereo audio. Pika’s hosted API is where you get video at 1080p, 2K and 4K, plus the audio stems and prompt enhancement. The card also describes Flex as a post-trained H3 model that makes video with synchronized audio in six sampling steps.
Who it is for. The card lists multi-shot storytelling, motion graphics, product films and reference-guided animation as good fits.
What’s real here: open weights, synced sound and heavy reference control together are a real step for open video. Keep the split in mind, though: 4K output and separate stems are hosted API features, and the ranking is Pika’s report at launch.
Sources
By Willa Ravenwire · The AI Bandwagon
