Image & Video

AI News You Missed - March 2026

From NVIDIA's speed-focused 88B model to uncensored 122B giants, March saw a surge in specialized AI releases.

Deep Dive

The AI landscape in March 2026 was defined by a surge of highly specialized, powerful models and tools. NVIDIA made a major play with its 88-billion-parameter GPT-OSS-Puzzle, emphasizing serious inference speed. Meanwhile, the frontier of large, uncensored models expanded with Dealignai's Nemotron-Cascade-2-30B and the massive 122-billion-parameter Qwen3.5-122B-A10B-Uncensored, pushing the boundaries of unrestricted conversational AI. Beyond raw scale, developers focused on niche expertise: Meituan's LongCat-Flash-Prover tackles formal mathematical proofs, FPHam's Regency-Aghast-27b writes in the style of Jane Austen, and OpenBMB's MiniCPM-o-4_5 handles real-time vision and voice.

This specialization extended across modalities. In video, NVIDIA's SANA-Video accelerates 2K AI video creation, while Fudan-FUXI's OmniVideo2-A14B enables omnidirectional generation. For images, new distillation and quantization methods like Z-Image-Distilled and Z-Image-SDNQ-uint4-svd-r32 dramatically speed up and optimize generation. A notable counter-trend is the rise of highly efficient small models, such as OrionLLM's 3-billion-parameter GRM2, which packs significant reasoning power into a compact package. The month also saw crucial infrastructure releases, including Unsloth's optimized GGUF files for coding models and new datasets like MoonshotAI's WorldVQA for testing AI memory.

Key Points
  • NVIDIA released GPT-OSS-Puzzle, an 88-billion-parameter model focused on unlocking serious inference speed.
  • Multiple uncensored giants emerged, including Dealignai's 30B Nemotron-Cascade-2 and a 122B Qwen3.5 variant, pushing conversational limits.
  • Specialized tools for video (NVIDIA's SANA-Video), math proofs (Meituan's LongCat), and efficient small models (OrionLLM's 3B GRM2) defined the month.

Why It Matters

The shift toward specialized, high-performance models means professionals can choose AI tools fine-tuned for specific tasks like coding, reasoning, or content creation.