40 / Sep 29 – Oct 5, 2025 · 6 RELEASES

AI News Week 40, 2025

Week 40 2025 AI releases.

Language Models (4)

Step3-VL-10B
StepFun · Reasoning
A 10B parameter open-source multimodal vision-language model achieving frontier performance on STEM reasoning, visual perception, and OCR tasks through high-quality pretraining on 1.2T tokens and scaled multimodal reinforcement learning. Demonstrates competitive or superior performance against models 10-20x larger while maintaining efficiency.
Sarvam 30B/105B
Sarvam AI · Multilingual
Large language models with 30B and 105B parameters, likely optimized for Indian languages and use cases based on Sarvam's focus.
NVIDIA Nemotron-3 Super 120B
NVIDIA · MoE Architecture
A 120B parameter language model optimized for inference efficiency with a 12B active parameter configuration using mixture-of-experts architecture. Designed for high-performance deployment with BF16 precision support.
TADA-1B
Hume AI · Small Language Models
A 1 billion parameter model designed for multimodal understanding and generation tasks. Optimized for efficient deployment with focus on emotional and contextual awareness capabilities.

Qwen3-VL-Embedding-2B
Alibaba Qwen · Visual Understanding
A 2 billion parameter vision-language embedding model designed for multimodal understanding and retrieval tasks. Optimized for efficient deployment while maintaining strong performance on vision-language downstream applications.

Qwen3-VL-Reranker-8B
Alibaba (Qwen) · AI Infrastructure
An 8B parameter vision-language reranker model designed to improve retrieval and ranking tasks by combining visual and textual understanding. Optimized for efficient multi-modal ranking of search results and document retrieval scenarios.

Published by Quanta AI Labs, an AI agency in India. Browse every week in AI News Weekly.