Open-weight video and multimodal models undercut frontier prices, but provenance lags

The cost case for open models

For buyers, this is a spending argument first. A team once paid frontier prices for every task. It can now reserve the expensive model for the hard tenth. It can run the rest on weights it can download. Smaller multimodal models also run on devices, not only in a vendor's cloud. That cuts latency and data movement. The cheaper stack does not fix provenance. Open weights are easy to copy, rename and pass off as something else. No common verification standard exists. Regulators and platforms still have no simple way to tie an output to a specific model.

The catch is the tenth that stays on the frontier model. Someone still has to decide which tasks qualify. A wrong call is cheaper to make than to fix.

Video and embeddings are two ends of the same shift. Capability that used to sit behind an API now ships as weights.

Pika released H3 Flex on 29 October. It is an open-weight video generation model for professional creators, developers and researchers. The company post-trained MiniMax H3 to generate video in six steps instead of 50. It then refined the model for visual quality and creative control. Model adapters sit on Hugging Face. The inference code is on GitHub. Pika pitches the model as a tool for iteration, not one-shot generation. That is where the saved seconds add up.

Pika puts video generation in the open

Pika H3 Flex handles text-to-video, image-to-video and reference-based generation.

Creators can supply up to 12 reference files across images, videos and audio. In Pika's example, one image fixes a character. A photograph sets the environment. A clip shows the intended camera move. An audio file carries the sound. Native output is 480p and 768p. Pika's hosted services extend it to 1080p, 2K and 4K. Generated video carries synchronised stereo audio. The hosted API exposes five separate audio stems.

Pika claims a 15-second 768p clip with audio takes about seven seconds of transformer computation under a controlled benchmark. That is computation time, not end-to-end time. Reference processing, decoding, service overhead and higher-resolution finishing all add to the wait.

Fewer steps shorten the loop between attempts. Creators can test prompts, camera moves and characters with less waiting between iterations.

Smaller multimodal models reach devices

Google's EmbeddingGemma team went the other way on size. Its 744M-parameter EmbeddingGemma 2 encodes text, images, audio, video and combinations of them into one 768-dimension embedding space. The budget splits across a 271M text backbone, a 167M vision encoder and a 305M audio encoder. On-device numbers come from a Pixel 11 Pro. The text-only Q4_C quantized variant peaks at 191 MB there. The full multimodal stack takes 567 MB. Native audio retrieval scores 69.54 MRR@10. That compares with 50.95 for Google's earlier EG1 ASR-plus-text pipeline. The paper claims the model beats specialist systems and models two to twelve times larger on multimodal benchmarks.

Matryoshka Representation Learning lets callers truncate the embedding to 128 dimensions. Text-only shops can skip the other towers.

Size buys portability, not reasoning. The published scores are retrieval benchmarks, not generation tests.

Verification and safety lag the releases

Anthropic is building an open source system. It will prove which model produced an output.

A blog post published on 3 November sets out the system, FlashProve. It would require labs to publish a fingerprint of each model. Labs can later match that fingerprint to outputs and verify it cryptographically. The fingerprints do not reveal model weights. Commercial labs could publish them without giving up their most valuable asset. Anthropic published the specification and working code on GitHub. It also said there are no plans to deploy the system at Anthropic or any other lab.

A University of Hong Kong study puts numbers on the safety gap. ByteDance's Seedance 2.0 Pro topped the study's visual quality ranking at 3.84 out of 5. Google DeepMind's Veo 3.1 scored 3.47. OpenAI's Sora 2 Pro scored 3.27. Seedance also led on visual consistency, physical realism, aesthetic quality and instruction compliance. Seedance did less well on safety. The same study placed it sixth out of 11 models for blocking harmful content. Zhenhui Jack Jiang led the research. He is a professor of innovation and information management at the university's business school.

Anthropic describes the FlashProve work as intermediate progress. It asks others to take it and improve it. No lab has committed to publishing fingerprints. No deployment date exists.