October’s open-source releases split between large models that handle many kinds of media and small specialists. Xiaomi’s MiMo-V2.6 is a family that handles text, audio and video, led by a 1.02-trillion-parameter model that activates 42 billion per token, alongside a 310-billion-parameter Flash version, and Xiaomi released the reinforcement-learning environments it trained on, spanning code, cyber, general, visual and music tasks. Kandinsky 6.0 Video generates five-second clips with synchronized 44 kHz audio, lip-sync included, from 3-billion and 29-billion-parameter models that upscale to Full HD. YuE2 writes a readable score before rendering it as a full song, and expert listeners preferred that planning step by 49.3% to 34.6%.
The smaller releases show where open models are heading: narrow, cheap jobs. Google’s EmbeddingGemma 2 is a 740-million-parameter model that turns text, images, audio and video into one common set of numbers, so a single search can match across all four, and it runs on a phone in about 567 megabytes. PolicyLM-1.7B is a moderation classifier that reads a platform’s own rules with every message, returns a score from 0 to 1 for each category, and answers in under 100 milliseconds, with no retraining when the rules change.
Open models are dividing the work. The largest compete on scale, and increasingly ship the training recipe, not just the weights. Small models take the narrow jobs, small enough to run on a phone or a single graphics card.
Read More: the previous open-source radar, on an open code corpus and audits that check themselves
Sources:
- MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement (arXiv)
- Kandinsky 6.0 Video: Foundation Models for Synchronized Video and Audio Generation (arXiv)
- YuE2: Unifying Symbolic and Audio Music Generation at Frontier Quality (arXiv)
- EmbeddingGemma 2: an open, lightweight multimodal embedding model (Google)
- EmbeddingGemma 2 model card (Hugging Face)
- Introducing PolicyLM-1.7B (Musubi)
Disclaimer: For information only. Accuracy or completeness not guaranteed. Illegal use prohibited. Not professional advice or solicitation. Read more: /terms-of-service
Reuse
Citation
@misc{kabui2026,
author = {{Kabui, Charles}},
title = {Open {Source} {This} {Month:} {Video} {With} {Sound,} a
{Trillion-Parameter} {Model,} and a {Moderator} {That} {Reads}
{Your} {Rules}},
date = {2026-10-13},
url = {https://toknow.ai/posts/open-source-radar-mimo-v26-kandinsky-6-embeddinggemma-2/},
langid = {en-GB}
}