Is Pika H3 Flex state of the art?Pending benchmarks.Announced Oct 29, 2026; no independent results yet.
Pika's video-with-audio model, post-trained from MiniMax's MiniMax-H3 to generate in six sampling steps. It makes 5–15 second clips with synchronized stereo audio from text, first and last frames, or up to 12 image, video and audio references. Open weights (LoRA adapters on MiniMax-H3) under the MiniMax H3 Community License, plus the Pika API, which adds 1080p to 4K upscaling and separate audio stems.
The details so far
- Lab
- Pika
- Announced
- Oct 29, 2026
- API
- Live
- Open weights
- Published · Hugging Face (Pika-Labs/Pika-H3-Flex): two FP32 LoRA adapters, FL2VA and Ref2VA (r = alpha = 128), with Comfy-format BF16 LoRAs; the repo also bundles the MiniMax-H3 base components (about 210 GB)
- Base model
- MiniMaxAI/MiniMax-H3, post-trained by Pika (a derivative; not a MiniMax release)
- License
- MiniMax H3 Community License Agreement for the adapters (card license: other); example code Apache-2.0
- Modes
- Text to video and audio; first-frame, last-frame and first-and-last-frame conditioning; multimodal references
- Inputs
- Prompt of up to 8,000 characters plus up to 9 images, 3 videos and 3 audio clips (12 references max; at least one image or video for reference mode)
- Output (open weights)
- Native 480p or 768p in 11 aspect ratios, 5–15 seconds at 24 fps, with 32 kHz stereo audio mixed into an MP4
- Sampling
- Six generator evaluations with the released adapter's schedule
- Pika API extras
- Upscaling a draft to 1080p, 2K or 4K; five separate audio stems (dialogue, score, SFX, Foley, ambience); prompt enhancement. These are hosted-only and not part of the weights
- Runtimes
- Pika's inference package (one request per GPU; tested on H100 80GB), Hugging Face Diffusers and ComfyUI workflows
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: #1 among open-weights models and #7 overall on DesignArena's text-to-video leaderboard at launch.
Lab claim: a 15-second 768p clip with audio in as little as seven seconds of transformer computation (not end-to-end API latency).
Sources
LABPika blog: Introducing Pika H3 Flexpika.art/blog/pika-h3-flex · Oct 29, 2026LABPika announcement on Xx.com/pika_labs/status/2115860471215788262 · Oct 29, 2026REPOPika-Labs/Pika-H3-Flex on Hugging Facehuggingface.co/Pika-Labs/Pika-H3-Flex · Oct 29, 2026REPOPika-Labs/Pika-H3-Flex on GitHubgithub.com/Pika-Labs/Pika-H3-Flex · Oct 29, 2026Tracking continues until an independent board lists it.