DeepSeek V3.2-Exp

Chinese AI firm DeepSeek unveiled DeepSeek-V3.2-Exp, an “intermediate” model building toward their next architecture. It introduces Sparse Attention mechanisms to reduce compute cost and improve efficiency.
They also cut API pricing by approx. 50%, making it more accessible.
Meta “Vibes” AI-generated short video feed

Meta launched Vibes, a feed of short-form AI-generated videos (accessible via meta.ai / the AI app). Users can browse, remix content, share, etc.
Adobe Firefly Boards (global launch)

Adobe expanded Firefly Boards to global availability. It’s an AI-first collaborative space emphasizing video models and media creation features.
Google DeepMind’s robotics AI models (Gemini Robotics 1.5 / ER 1.5)

DeepMind introduced upgraded robotics models that can handle tasks like sorting laundry and classifying recyclables, combining vision with action and with planning.
Meta Ray-Ban Display (AI 2025 glasses with visual output)

At Meta Connect, they revealed Ray-Ban Meta Display glasses with a full-color display built into the lens with a Neural Band (gesture control via EMG) for interactions.
NVIDIA + Intel collaboration on AI infrastructure & AI SOCs

NVIDIA and Intel announced a joint project to build custom x86 chips integrated with NVIDIA’s GPU/AI architecture (RTX SOCs) to support cutting-edge AI workloads on client devices and data centers.
Moonshot AI Kimi K2 update

Moonshot AI, which had earlier released K2, rolled out Kimi-K2-Instruct-0905 in September 2025. The update doubled the context window (to 256K tokens) and improved performance in coding/agentic tasks.
They also released Kimi-K2-Instruct (0905 version), enhancing model capabilities.
Qwen model lineup expansions

Alibaba’s Qwen family saw multiple releases in September:
Qwen3-Max – a more capable “non-reasoning” foundation model.
Qwen3-Next – new architecture involving hybrid attention, sparse MoE structures, with efficiency and throughput improvements.
Qwen3-Omni – a multimodal model (text, image, audio, video) with streaming output capabilities.