AI News Roundup: Pyramid Flow, Video Input LLM, Gemini 2.0 & more!

Open-source video generation just took a major step toward “single-GPU fine-tuning,” with a new repository of memory-optimized training scripts aimed at the Cog family of models. The pitch is straightforward: a 5 billion-parameter video model should be tunable using a 24 GB GPU—far more accessible than the multi-GPU setups many teams rely on. Cog video x Factory packages that capability into a practical workflow, lowering the barrier for developers who want to adapt open models for specific styles or tasks. The bigger implication is modifiability: if fine-tuning becomes routine, creators can target niche outputs like animation, specialized upscaling, or domain-specific footage rather than settling for generic text-to-video results.

That push toward control and customization shows up again in Runway’s Gen 3 Alpha turbo update, which adds image input for both endpoints of a generated clip. Instead of using an uploaded image as only the first (or last) frame, users can choose one image for the start and another for the end, letting the model generate the in-between frames. When the two images differ only slightly, the feature can increase continuity; when they differ more, it enables deliberate transitions. Runway also positions Gen 3 as a speed-and-reliability leader among video generators, making it a practical option for longer projects.

The most headline-grabbing development is Pyramid Flow, a fully open-source text-to-image-to-video model released under an MIT license. Built around training-efficient autoregressive video generation using flow matching, it’s trained on open datasets and targets high-quality outputs at 24 frames per second. Reported resolution is a little over 720p, and the model checkpoints are available on Hugging Face, setting up a fast path for community installs and experimentation. Early demos emphasize smooth motion and convincing environment behavior—waves hitting rocks, camera pans across an ocean, and other landscape-style scenes. The results are described as strong for open source, though not always matching the very top closed models; still, the availability of checkpoints and training details is framed as a catalyst for rapid iteration, including future fine-tuning for specific creative needs.

On the multimodal language-model front, a new model from Rhymes AI (Apache 2.0–licensed) is drawing attention for taking both images and video inputs into a text-based system. The transcript highlights that this kind of video intake remains rare across mainstream chat platforms, where video upload is typically unsupported. The model is also positioned as fine-tunable, with examples including debugging a code screenshot and converting handwriting into text.

Meanwhile, Google Gemini 2.0 is said to be in the works, with hints from a DeepMind event pointing to a multi-turn successor to Frontier models, plus vision and likely audio-related capabilities. OpenAI’s ChatGPT also gets incremental interface and command changes, including web search returning via slash commands, while an “01 preview” reasoning command still can’t actually see uploaded images.

Outside software, Tesla’s humanoid robots appear on stage performing tasks like pouring drinks and playing Rock Paper Scissors, but the transcript repeatedly flags that teleoperation is likely involved. Finally, Meta AI adds a new voice mode with cloned voices, though it’s described as a pipeline (speech-to-text, response, then text-to-speech) rather than the more natural “native multimodal” voice approach associated with advanced voice systems elsewhere.