Pika H3 Flex open weights: 6-step video with synced audio
Two items this afternoon: an open-weights video model that writes its own soundtrack, and a quiet tightening of who gets to stack ChatGPT plans.
Pika has released Pika H3 Flex, an open-weights derivative of MiniMax's H3 video model that generates picture and synchronized stereo sound in a single six-step pass. The adapter weights and the inference code are public — code under Apache 2.0, weights under MiniMax's H3 community licence — and the package is a smaller deal than a frontier launch but a more useful one, because sound has been the missing half of every open video model shipped this year. Two LoRA adapters come with it: one handles text-to-video plus first-and-last-frame conditioning, the other takes multimodal references, so a workflow can start from a prompt, a pair of keyframes, or up to twelve reference assets — nine images, three videos and three audio clips — according to the model card.
What actually comes out the other end is 480p or 768p native, clips of 5 to 15 seconds in whole-second durations, encoded as H.264 with 32 kHz stereo AAC audio. The extras — 1080p, 2K and 4K upscaling, plus separation into five audio stems (dialogue, score, sound effects, foley and ambience) — run only on Pika's hosted API, not on your machine. That split is deliberate: generate drafts locally at low resolution, keep the take you want, then pay Pika to finish it. For a studio workflow it is a sensible division of labour; for anyone hoping to run the whole thing offline, it means the local model is a preview engine, not the product.
The interesting part is how Pika got from 50 denoising steps to six. The team fine-tuned MiniMax H3 on an in-house audio-visual dataset, then applied projected distribution matching distillation to compress the sampling trajectory, then trained preference-feedback and direct-preference LoRAs to repair aesthetics, sharpness and instruction-following — with the adapters merged at inference so the step count stays at six, per Pika's technical write-up. The hosted stack got a second pass: on eight NVIDIA B300 GPUs, a five-second 768p clip with audio went from 3.409 seconds to 2.129 seconds of total inference, with the checkpoint and the six-step schedule held fixed — a 1.60x speedup overall, 1.74x on the transformer forwards alone.
Treat that number with the care it needs. It is an internal measurement on a datacenter cluster with the same weights, not an API latency figure, and Pika's separate marketing line about seven seconds for a fifteen-second clip measures denoising compute only — the two should not be blended into one "Pika is faster now" claim. The honest reading is narrower and still useful: a six-step sampler on B300s puts a draftable 768p clip inside the two-second range, which is what makes the generate-review-regenerate loop tolerable.
On quality, H3 Flex sits at eighth on the DesignArena video leaderboard as of November 1, with an Elo of 1,288 across 817 battles and a 59.5 percent win rate — down one place from the seventh Pika reported at launch, when the score was 1,293. Eighth overall still makes it the highest-placed open-weight video model on the board: Wan 3.0 sits at fourth but ships no open weights, and MiniMax's own H3 is eleventh. Pika's internal comparison is less flattering than the headline. Across its own seven audio-visual metrics, H3 Flex wins five — and loses audio-visual alignment to fal's H3 Max, 0.212 against 0.216, and text-video alignment, 0.225 against 0.226. All of those are Pika's numbers about Pika's model, which is the standard arrangement for a launch and the reason independent boards matter more than launch tables.
The wider context is that MiniMax H3 has become the base layer of open video. Released in August, it has since spawned variants from Running Hub, fal, FastVideo and an NVIDIA–MIT team, and Hugging Face currently lists 472 derived repositories off the base — 173 adapters, 204 fine-tunes, 11 merges and 84 quantizations. MiniMax's claim of more than 24 million downloads in the model's first month is company-stated and cannot be confirmed from public download counters, which only expose 30-day figures. What the repository count does show is that the fine-tuning ecosystem is real and fast, and that a Western lab building on a Chinese base model is now an ordinary move rather than a controversial one.
The counter-argument to "open video with audio" as a milestone is the weight size. The base MiniMax H3 components run to roughly 210 GB before you add a single adapter, so this is open in the licence sense and not the practical sense; the people who can run it are the people who already have a B300 node or rent one by the hour. Openness here buys reproducibility and the right to fine-tune, not local generation on a gaming card. That distinction matters when the next round of "open model beats closed model" headlines arrives.
What to watch: whether an independent eval reproduces the DesignArena placement once the initial launch burst of battles settles, and whether anyone distils H3 Flex down to a consumer-GPU footprint — that, not another leaderboard point, is what would put audio-native video generation in front of people who are not studios.
OpenAI has put caps on how many paid ChatGPT subscriptions one payment card or one device can carry. The company's help centre now states, without giving a number, that if too many subscriptions are associated with the same card or device, ChatGPT may block a new purchase — with the message "Too many subscriptions associated with this card or device," which OpenAI distinguishes from a bank decline. A separate device cap refuses sign-in with "Device limit reached" per the help centre article.
The limits cover ChatGPT Go, Plus, Pro 100, Pro 200 and Pro 500. Free accounts and the Business, Enterprise and Edu workspace plans do not count toward them. Three details do more work than the headline: signing out of an account does not guarantee another paid account becomes eligible, buying credits or resetting usage will not clear a card or device restriction, and subscriptions bought before the change keep working under an existing-subscription exception — but cancelling one does not guarantee you can replace it. OpenAI frames the whole thing as privacy and security protection, and offers cancellation-and-refund for subscriptions caught by it.
There is no announcement, no effective date and no published number, so what we can say is bounded: the restrictions are documented and live, the thresholds are not. The obvious population affected is the one that stacks cheap flat-rate plans on one device to get more usage than a single subscription allows; OpenAI's suggested alternatives — extra credits, usage resets — cost more per unit than another subscription did. Whether that is revenue discipline or fraud control is a judgment the page does not settle. One clarification worth keeping straight: the widely repeated "two accounts per device" figure is a different limit, on simultaneous sign-ins, and the help page says explicitly that device limits are separate from account switching.
Do these caps make the flat-rate subscription model unsustainable, or are they just closing an abuse path? Write to us via our contact page.
