Open-Source AI Boom: IBM’s Granite 4.0 & Qwen3-VL Lead Wave of New Models
New Open-Source AI Models & Releases
October 2025 Open-Source AI Model Roundup: A comprehensive list of newly released open-source AI models spanning audio, vision-language, multimodal, and agent-based systems. Highlights include IBM’s Granite 4.0 (hybrid-efficient models), Qwen3-VL-30B/32B (advanced vision-language models), HunyuanWorld 1.0 (Tencent’s 3D world model), and LFM2 (LiquidAI’s multimodal series). The list also features specialized models like KANI-TTS (text-to-speech), DeepSeek-OCR, and StreamDiffusion (real-time generation).
- List of interesting open-source models released this month
- Full list with links: GitHub/HuggingFace repositories, IBM Granite 4.0, Qwen3-VL, HunyuanWorld, etc.
- ArXiv papers for technical details.
Vision-Language & Multimodal Models
Qwen3-VL Impresses with Advanced Image Analysis: Users highlight Qwen3-VL’s ability to draw bounding boxes, parse complex visual data, and generate detailed descriptions. The model’s instruct-tuned variants (30B/32B) are noted for robustness in real-world applications like document understanding and scene analysis.
Developer Tools & Integrations
glm-proxy Fixes GLM 4.5 Air’s Tool Call Issues: A new open-source proxy server converts GLM 4.5 Air’s XML-formatted tool calls into OpenAI-compatible JSON, enabling seamless integration with existing OpenAI clients. The tool addresses a key interoperability gap for developers using GLM models.
Notable Mentions (Standalone Releases)
OpenAI Releases GPT-OSS Safeguard: OpenAI introduces a new open-source safety toolkit ("GPT-OSS Safeguard") to help developers implement alignment and moderation layers in custom AI models. Details are sparse, but the announcement signals a push for standardized safety practices in open-source AI.
- OpenAI’s announcement (linked in the original post).
Minimax Unveils M2 Model: Minimax announces M2, a next-gen multimodal model with claimed improvements in reasoning and efficiency. The release targets enterprise applications, though technical specifics are pending.
NVIDIA Nemotron-Nano 12B-V2-VL: A compact but powerful vision-language model optimized for edge devices, now available on NVIDIA’s model hub. Designed for low-latency applications like robotics and AR/VR.