Pika Labs has open-sourced Pika H3 Flex, a post-trained take on MiniMax's H3 video model that cuts sampling from fifty steps to six while keeping, in Pika's words, quality intact. The code is public and free to run, and the release includes two trained adapters covering the workflows creators actually use.
The training recipe is straightforward: the team improved aesthetics, clarity and instruction following through reward feedback learning and direct preference optimization, training LoRA adapters rather than touching the base model. At inference the adapters merge back in and generation stays at six steps. The first open adapter handles text-to-video plus first-and-last-frame conditioning; the second handles multimodal reference generation, where creators combine images, video clips and audio to steer a character's look, a scene, motion and camera moves.
Output is native 480P and 768P, with clips of roughly five to fifteen seconds that carry synchronized stereo audio — a notable detail, since most open video models still ship silent. The release ships ComfyUI workflows and HuggingFace Diffusers integration, so the whole pipeline runs locally for anyone with the hardware.
Pika's comparison numbers are self-reported and not independently verified: in the company's same-prompt, 768P evaluation, H3Flex scored highest among compared models on five visual and audio quality metrics, and Pika says it sits eighth on the text-to-video Arena, the highest placement of any open model on that list. Treat those as the vendor's own scoreboard until a third party reruns them.
What is not self-reported is the ecosystem math. MiniMax open-sourced H3 three months ago, and it has become one of 2026's most downloaded open models — over 24 million downloads in its first month and more than 300 public derivative models. FastVideo shipped FastH3, a four-step distilled version; fal built H3Max inference optimizations; and last week NVIDIA and MIT released Open-H3-Regenerate-2K, a 2K/4K video regeneration stack on the same base. Pika's post-training adds a creator-facing layer to that stack.
The hosted side stays commercial and separate. Pika's API offers 1080P, 2K and 4K enhancement plus audio separation into five stems — dialogue, music, sound effects, foley and ambience — so a creator can generate a low-resolution draft locally, pick the takes, then pay for high-resolution finishing and a split-track mix.
The strategic read: open base models are collapsing the distance between research releases and production tools. Three months after H3's weights landed, a consumer video company has rebuilt its core generation loop on top of them and given the results away. For MiniMax, every adapter like this raises the value of the base model it gave away. For closed-lab video models, the moat is shrinking to what the open stack cannot yet do — resolution, length and reliability at scale.
Comments (0)
Log in to join the discussion
Log InNo comments yet