Slop TVNewsLatest
News

Pika H3 Flex cuts video generation from 50 steps to six

Pika's post-trained MiniMax H3 turns out a 15-second 768p clip in as little as seven seconds of denoising time, and the API starts at a cent a second.

Illustration: Pika H3 Flex cuts video generation from 50 steps to six
Illustration: SLOP TV News. Logos are trademarks of their owners.

Key takeaways

  • Pika launched Pika H3 Flex on October 29, 2026, a six-step video model post-trained on MiniMax H3 that generates synchronized audio alongside the picture.
  • Pika's own research lead, Cade Li, says the model was brought down from a 50-step teacher to six sampling steps and then refined for visual quality and creative control.
  • Local inference runs one request per GPU and was tested on an H100 80GB, with roughly 215 GB of files to download for the released weights.
  • Only 480p and 768p run locally; 1080p, 2K and 4K finishing and the five audio stems are Pika-hosted features.

Pika shipped Pika H3 Flex on October 29, 2026, a version of MiniMax's H3 video model that the company post-trained to generate picture and synchronized audio in six sampling steps instead of 50. Pika's launch post says the model is the highest-ranked open-weights entry and seventh overall on DesignArena's text-to-video leaderboard, and the launch release puts a 15-second clip with synchronized audio at 768p at as little as seven seconds of denoising time. The release footnotes that figure as denoising time, not end-to-end API latency.

Pika, the Lightspeed-backed San Francisco studio founded in 2023 and known for Pikaffects, framed the release around iteration, per the launch release. Creators draft at 480p or 768p and finish at 1080p, 2K or 4K, hand the model up to 12 references, and export five audio stems. "We brought generation down from 50 steps to just six," research lead Cade Li said in the release.

The released weights are Model Derivatives of MiniMax H3. They ship under the MiniMax H3 Community License Agreement, not an open-source license. Its Applicable Territory is the world minus four places: the European Union, the United Kingdom, the Republic of Korea and the United States. Section V, item 4, of the use restrictions says that using the works or their outputs outside that territory is not authorized. MiniMax claims no rights over the outputs you generate and leaves them your responsibility, but bars using them to improve another AI model and asks for separate written authorization above $20 million in yearly revenue. Pika's own pages do not flag the exclusion.

Independent write-ups split on what to make of it. AI Weekly calls the 33B-parameter model a deliberate split, with open weights for draft resolution and the API for anything a client would ship. Neural Brew notes it had not verified the leaderboard placement, and AlphaSignal lists the Elo at 1293, ahead of MiniMax H3 and H3 Max at 1277.

Local inference takes one GPU per request, tested on an H100 80GB, with about 215 GB of downloads. Pika's API prices H3 Flex from $0.01 a second at 480p, $0.08 at 4K and $0.001 a second for stems, and fal lists H3 Max from $0.03 a second. Pika bills its model as 1.8x faster than fal's H3 Max and up to 80% cheaper.

Pika H3 Flex weights are on GitHub and Hugging Face now; the API is live, and the creation product gets it on November 2, with 15 free generations a day and 50% off for members and subscribers for two weeks.

Sources

  1. pika.art - launch post: date, six steps, 12 references, five stems, 4K, open weights, November 2
  2. pika.art - technical deep dive: six-step schedule, latency, per-second pricing table
  3. businesswire.com - release: quotes, DesignArena ranking, seven-second footnote, launch promotion
  4. huggingface.co - model card: specs, hardware, license, 12-reference limits
  5. huggingface.co - MiniMax H3 Community License text: territory, outputs, revenue threshold
  6. dev.pika.art - Pika API endpoint pricing for H3 Flex
  7. fal.ai - fal's listed per-second rates for H3 Max
  8. fal.ai - fal's own MiniMax H3 and H3 Max per-second rates
  9. aiweekly.co - 33B parameter count, download size, "splitting the stack" reading
  10. neuralbrew.ai - license and hardware limits, ranking marked unverified
  11. alphasignal.ai - Elo 1293 leaderboard table, six-step schedule, launch pricing table
  12. comfyui-wiki.com - ComfyUI LoRA packaging and templates