Multimodal

Google's open 740M-parameter multimodal embedding model runs on-device for image, video, and voice search. A local demo indexed 300 photos and reached 82.5% precision at 10 on held-out animal queries.

Runway Labs previews Project Continuum, a research app built around real-time video interfaces. Demos show animated browser overlays and scenes that recompose when resized.

Meta's VR glasses move compute and battery to a wired pocket module, bringing headset weight to about 100 grams. Reports cite a 5K micro-OLED display and spring 2027 availability.

Google introduced Gemini 3.8 Live and Live Extended Thinking in Gemini Live and its API. The models add near-real-time visual understanding, support for 97 languages, and background tool use.

Runway published a research overview and demos describing media that adapts to user input in real time, including a user-steered simulation with AI. The company says instant generation is the direction for video models after its Solaris and GWM Worlds 2 releases.

Runway says GWM Worlds 2 generates continuous interactive 720p video at 24 fps with 48 kHz audio. WorldPrompt separates persistent world rules from changing actions and events.

Google’s Agentic Video processing lets Gemini reason across a video’s transcript, audio, and frames rather than scan the whole file uniformly. Google says dynamic frame sampling can cut token use by up to 88% on supported tasks.

Runway introduced Solaris, an Interface World Model that generates interactive interfaces frame by frame without code. The research is not yet publicly available.

DIT AI says its OpenAI-compatible API routes text, image, and video calls among more than 50 models and 160 providers. The company claims typical costs 30–70% below official API pricing; that savings figure is vendor-reported.

MUZIM presents Vibe Search as locating a scene at a specific timestamp from a text description rather than returning only a matching video. The tool also searches and organizes files across a user’s existing library.

DeepSeek's experimental V4 Flash Vision model is reportedly live on its official API. In a screenshot-only form-filling test, it completed and checked fields in 5 minutes 48 seconds but misplaced check marks and drifted in longer boxes.

Pika Music accepts text, lyrics, voice, and music references separately or together in one diffusion decoder. It is available through Pika API Club, and Pika claims up to 10 times better cost efficiency than other music models.

Alibaba released Qwen3.8-Max, described in a launch thread as a 2.4T sparse MoE with 95B active parameters, 1M context, native vision/text, and agent benchmarks. API pricing is listed at $2/$6 per 1M tokens, with open weights planned for Hugging Face.

Practitioners said coding agents can make impressive game demos quickly, but playable games still require weeks or months of design, testing, and cleanup. Other posts pointed to limited visual reasoning, 3D topology fixes, and Fable or Opus model choices as current bottlenecks or workarounds.

Venturetwins built a podcast clipping app with Thinking Machines’ Inkling. The app analyzes long-form audio, selects clip candidates by topic or best moment, and directs FFmpeg edits.

Google DeepMind unveiled GenCeption, which turns video into depth, segmentation, camera rays, 3D keypoints, and prompt-steered scene representations. The thread links project and Hugging Face materials.

Goodside compared Claude Fable 5 Max puzzle generation with GPT-5.6 Sol Pro solving attempts. Sol solved a 150-Pokémon empty crossword but failed the 1,025-Pokémon version, with one reported success traced to the answer key.

Goodside tested GPT-5.6 Sol and Claude Fable 5 with fake handwriting, constrained-vocabulary prompts, and Ghost Font. Sol often answered unreadable inputs, while Fable more often refused or pushed back.

Goodside fed GPT-5.6 Sol and Claude Fable 5 binary noise and meaningless handwriting. Sol often invented hidden text, while Fable also failed on noise but more often pushed back on scribbles.

Levelsio exported diet logs from Claude chat into Claude Code on a VPS, then built Caltrack with Telegram, SSH entry, database memory, and Grok vision food scanning. The case turned failed chat memory into a persistent app.

OpenAI staff said the new ChatGPT Voice is powered by GPT-Live-1 for more natural conversations. The launch also refreshed the ChatGPT shader with Blender prototypes translated to Metal and WebGL using Codex.

Posts say Meta AI began rolling out Muse Image and Muse Video in the U.S. Muse Image is described with web search, code execution, self-critique, multi-reference composition, and Content Seal watermarking.

xAI opened Grok Imagine 1.5 Preview in its Imagine API, moving the model from benchmark chatter into direct creator access. The same-day Cloudflare AI Gateway support gives teams another route to run Grok models in production.

Google’s new Gemma 4 12B ships as an encoder-free open model for text, image, audio, and video tasks with a 256K context window. Early GGUF ports and local benchmarks make it a plausible on-device multimodal option for creator tooling and experimentation.

A sponsored explainer thread described Speech Engine as a WebSocket layer that adds speech-to-text, turn detection, interruption handling, and text-to-speech to existing LLM agents. The pitch is that teams can keep their current model stack and add voice without rebuilding the whole agent.

Posts report SenseTime open-sourced SenseNova U1, a unified text-image model with interleaved generation, 8-step distilled LoRA and ComfyUI workflows. They cite 2K image times around 15 seconds and H100 inference cuts to about 2 seconds, so compare it against your current image pipeline.

Tencent released HY-World 2.0 with WorldMirror 2.0 code and weights for turning text, images, or video into persistent 3D scenes. The output includes navigable geometry and camera data instead of disposable video frames.