MiniMax H3, an open‑source multimodal video model, achieves SOTA performance, supports text, image and video inputs with native stereo audio, costs as low as 0.5 CNY per second, and demonstrates seamless editing capabilities that rival traditional After Effects workflows.
An eight‑year‑old Indian developer named Laksh created a fully offline, voice‑controlled AI storytelling device using an $8 ESP32‑S3 microcontroller, an INMP441 microphone, a MAX98357A amplifier, and the open‑source TinyStories LLM, demonstrating that small language models can run without cloud connectivity.
Independent developer David Ondrej released SuperGemma 26B, an uncensored fork of Google’s Gemma 4 26B that removes all safety guardrails, runs locally on consumer‑grade GPUs, and has sparked intense debate over its technical merits, deployment simplicity, and the security risks of a truly unrestricted AI model.
Semgrep’s benchmark shows that the open‑source GLM 5.2 model, using only a unified prompt and a lightweight Pydantic AI scheduler, achieves a 39% F1 score on IDOR vulnerability detection—outperforming Claude Code’s best 37.4% while costing only about $0.17 per discovered flaw.
JD’s open‑source JoyAI‑VL‑Interaction model transforms large‑language models from static question‑answering to continuous visual‑language interaction, enabling proactive judgment, instant responses, and intelligent task delegation, with benchmark win rates up to 87.9% against leading competitors and full stack code, model, and dataset released for real‑world deployment.
Google’s Gemma 4 12B delivers near‑26B performance with half the memory, runs on a 16 GB laptop GPU, and uses a novel encoder‑free unified architecture that natively handles vision, audio, and text, making high‑quality multimodal AI truly local.
SenseNova U1 Lite, an 8‑billion‑parameter open‑source multimodal model from SenseTime, uses the NEO‑Unify architecture to fuse vision and language in a single space, achieving commercial‑grade efficiency and benchmark scores that surpass much larger proprietary models while supporting continuous image‑text generation.
JoyAI-Image-Edit, an open‑source multimodal foundation model from JD Research Institute, integrates text‑to‑image generation, image understanding, and instruction‑driven spatial editing, achieving world‑leading spatial perception and editing capabilities that unlock new applications across e‑commerce, robotics, 3D reconstruction, and design.
Kimi K2.5, Moonshot’s latest open‑source multimodal model trained on 15 trillion image‑text tokens, adds native vision capabilities and a 100‑agent swarm that speeds complex tasks by 4.5×, achieves top‑tier benchmark scores, and can be deployed with vLLM, while demanding significant resources and hardware.
Tencent has open‑sourced its upgraded 1.5‑billion‑parameter HunYuan DiT model—the first Chinese‑native, bilingual (Chinese‑English) text‑to‑image diffusion‑with‑transformer system—delivering about 20% visual quality improvement, multi‑round generation, video‑generation potential, and free commercial use, with full weights, inference code, and algorithms available on Hugging Face and GitHub for developers and enterprises.
