/
Published on 19.03.2026
TLDR: OpenAI launched two new compact models — GPT-5.4 mini and nano — that deliver near-flagship performance in smaller, faster packages optimized for coding and high-volume API workloads. Meanwhile, open-source Mamba-3 is proving that Transformer architectures aren't the only path forward, outperforming them on key benchmarks while running significantly faster.
OpenAI's latest move in the compact model space is significant. GPT-5.4 mini and nano are designed specifically for coding tasks and high-volume API usage, hitting a sweet spot between performance and cost that most teams building AI-powered products actually need. The industry has been chasing larger and larger models, but the real engineering value often lives in models that run fast, cost less, and still deliver strong results. These two models aim squarely at that gap — near-flagship capability without the flagship price tag or latency.
The more disruptive story in this roundup is Mamba-3. This open-source model outperforms Transformers by 4 percent on language benchmarks while running up to 7 times faster on long sequences. That's not a marginal improvement — it's a fundamentally different architectural approach showing competitive results. The Transformer architecture has dominated since 2017, and most of the industry's infrastructure, tooling, and optimization work is built around it. Mamba's state-space approach challenges that assumption directly. If state-space models can match or beat Transformers on quality while being dramatically faster on long contexts, the implications for inference costs, real-time applications, and edge deployment are substantial.
On the regulatory front, U.S. senators are demanding ByteDance shut down Seedance 2.0 after it generated videos using real actors and licensed characters without permission. This is the kind of IP and consent issue that's been building since generative video became viable. The technology can now produce convincing video of real people and copyrighted characters, and the legal frameworks haven't caught up. This will be a defining tension in the generative video space for the foreseeable future.
Google is expanding Gemini's personalization features to free users across AI Mode in Search, the Gemini app, and Chrome. This is a direct play to make Gemini the default AI assistant across Google's ecosystem by lowering the barrier to personalized experiences. Meanwhile, Mistral launched Small 4, combining reasoning, multimodal vision, and agentic coding capabilities into a single configurable open model — continuing the European AI lab's strategy of shipping capable open-weight models that compete with closed offerings.
The roundup also highlights several emerging tools worth noting. Stitch from Google turns prompts, sketches, or screenshots into full UI designs and production-ready code in minutes. Paperclip is an open-source platform for building autonomous companies with AI agents. QuiverAI generates and animates production-ready SVGs from text prompts with a16z backing. And DittoDub translates and dubs videos into 36-plus languages using AI voice cloning that preserves the creator's original tone.