Weekly multimodal AI roundup: Phi-4, Helios, LTX-2.3

The pace of capable open-weight multimodal releases is accelerating fast enough that a weekly roundup is now genuinely hard to keep up with — and that's a bullish signal for the local AI ecosystem.

• –Microsoft's Phi-4-reasoning-vision-15B is the headline: MIT-licensed, 15B parameters, targeting math, science, and UI reasoning — a strong open-weight alternative to proprietary vision models
• –Helios (PKU-YuanGroup) claims 14B video generation running real-time on one GPU with t2v/i2v/v2v up to a minute; the author flags the numbers as suspiciously good, worth independent verification
• –LTX-2.3 from Lightricks shows healthy community momentum — GGUF workflows, a desktop app, and a Linux port emerged within days of release
• –NEO-unify skipping traditional encoders entirely is an architectural bet worth watching; growing evidence that CLIP/SigLIP encoders may not be essential for multimodal models
• –Tencent's HY-WU delivering face swaps and style transfer without any fine-tuning is a practical win for deployment scenarios where per-user training is infeasible