AI RADAR

NVIDIA — NVIDIA dropped its Nemotron 3 family of fully-open models, and the headline for solo builders is the Nano: a 4B model with GGUF weights on Hugging Face that runs right on your laptop, no cloud bill; the lineup scales up to a 550B Ultra reasoning model with a 1M-token context that went live day-zero on OpenRouter, Together, and Perplexity, and NVIDIA shipped the training data and recipes too under a permissive open license, so you can actually fine-tune instead of just downloading

OpenAI — ChatGPT's memory grew up — a new 'dreaming' architecture quietly revises what it knows about you in the background as your life changes ('you're going to Singapore in July' becomes 'you went'), so recommendations stay current instead of stale; paid users get double the memory capacity with summaries you can review and edit, and a 5x compute cut means real memory is finally rolling out to Free users over the coming weeks — your assistant stops forgetting who you are

Microsoft — Microsoft's first homegrown coding model lands inside GitHub Copilot for individual users — turn on auto-pick or grab it from the model picker in VS Code, no setup needed; it beats Claude Haiku 4.5 on coding benchmarks while burning up to 60% fewer tokens, and dials reasoning up or down per request so simple edits stay snappy; one of seven new MAI models (plus a reasoning model and a voice cloner) Microsoft shipped at Build to cut its OpenAI bill and pass the savings to devs

Microsoft — Microsoft open-sourced a way to test whether your AI agent actually behaves — write your rules in plain English ('never give medical advice', 'always cite a source') and ASSERT turns them into scored test cases, generating tricky scenarios and grading the output; MIT-licensed and framework-agnostic, it plugs into LangChain, CrewAI, OpenAI's Agents SDK and 100+ models via LiteLLM, so a solo dev shipping an agent gets real evals instead of vibes

MiniMax — open-weight frontier coding model with a 1M-token context and native image/video input — it can even operate a desktop; on SWE-Bench Pro it edges past GPT-5.5 and Gemini 3.1 Pro and lands just behind Opus 4.7, while the API runs $0.60/M in and $2.40/M out (cents on the dollar next to the closed labs); weights and the technical report hit Hugging Face within days, so you can self-host the whole thing

TwelveLabs — video-understanding startup TwelveLabs takes its tech straight to creators — describe the shot you want in plain English and Rodeo searches your entire footage library, finds the clips, and assembles a rough cut for you; powered by their Marengo 3.0 + Pegasus 1.5 models, it turns 'I know the clip is in here somewhere' into an instant edit instead of an afternoon of scrubbing timelines

Liquid AI — Liquid AI's new open-weight model runs a real tool-calling agent entirely on your laptop or phone — no cloud, no API bill, nothing leaving the device; it's a Mixture-of-Experts with 8.3B total params but only 1.5B active per token, so it stays fast and light enough for edge hardware while punching above its size on agentic tasks; weights are up on Hugging Face, making it a privacy-first base for any solo builder who wants assistants that work offline