Pika H3 Flex: 6-step open-weight video, 4K finish

Pika H3 Flex is a video model built for iteration speed: Pika took MiniMax H3, fine-tuned it on curated video, audio, and reference-based tasks, then distilled the 50-step teacher down to six sampling steps — and refined the faster model until it matched, and on several measured dimensions beat, the teacher [official]. The release is 33 billion parameters, shipped as LoRA adapters under the MiniMax-H3 community license [official].

It launched October 29, 2026 [official] — day-one coverage so far is press-release reprints plus Pika’s own blog, which is why this page is a status tracker, not a verdict. The interesting question is not whether 6 steps work in a benchmark, but what survives contact with real creators: local inference reports, the November 2 creation launch, and independent pricing.

Faster than real-time — with an asterisk. A 15-second 768p clip takes about 7 seconds of transformer (DiT) denoising in Pika’s controlled benchmark [official]. That is compute time, not the time from clicking generate to watching a clip: references, finishing passes, stem separation, and service overhead all add latency [official].

Draft fast, finish in 4K. The workflow Pika pushes is explicit: explore at 768p, then upscale only the version you love to 1080p, 2K, or 4K [official]. The upsampler preserves the generated content while adding detail [official] — but it is a hosted feature, not part of the open weights (see below).

Five audio stems. Export Dialogue, Score, SFX, Foley, and Ambience separately to refine any facet of the soundscape [official]. For creators who currently re-generate a whole clip to fix one sound, this is the most practical feature in the release — and, again, hosted-only.

Twelve references, precision edits. Create from up to 12 references — sketch, voice sample, image, sound, video, moodboard — and swap characters, camera moves, looks, and voices after the fact [official]. This is the “professional iteration” pitch: the model is built for revising, not for one-shot luck.

Benchmarks, honestly labeled. DesignArena lists Flex as open-weights #1 and overall #7 at launch [reported]. Pika’s own human-preference evaluation ranks it highest among tested models for anatomy, complex composition, audio sync, visual style, and camera control [official] — a vendor evaluation, not an independent one. No standardized benchmark compares it to Seedance or Kling yet [unknown].

“Open weights” here has a specific, limited meaning [official]. Open: generator adapters and inference code, text-to-video, first- and last-frame generation, multimodal reference-to-video, synchronized audio at 480p and 768p. Hosted-only: the 1080p/2K/4K upsampler, audio stem separation, optional prompt enhancement, and the accelerated serving stack.

So a local build gives you the draft engine; the finish — high resolution, separated stems, prompt help, fast serving — lives on Pika’s API. That split is the business model, and it is worth stating plainly: the open release is a funnel into the hosted API, where the per-generation economics actually sit. Nothing wrong with that, but “open” should not be read as “the whole product is free.”

No verified per-second or per-generation price exists yet [unknown]. What exists: Pika’s claim of 1.8x faster than fal H3 Max and up to 80% cheaper [official, vendor claim], and the Pika API Club launch offer — 15 free generations per day plus 50% off for the first two weeks [reported]. The creation product on pika.art opens November 2, 2026 [official].

What to watch: the November 2 launch (does the creation product match the API’s speed?), independent price reporting (does “80% cheaper” survive contact with real invoices?), and local inference reports (what hardware actually runs 768p at usable speed?). This page will update as those land.