Ant Group's Robbyant open-sourced a 1B vision model that beats models 7x its size at dense spatial perception. Tencent dropped a 295B MoE under Apache 2.0. NVIDIA unified audio and text in one decoder without losing IQ. And Liquid AI cut "doom loop" rates in reasoning models by 90% with one training trick.

Most vision foundation models optimize for semantic invariance — recognizing "this is a cat" regardless of pose or lighting — at the direct expense of dense spatial perception, the pixel-precise depth and boundary understanding that physical AI actually needs. LingBot-Vision flips the priority: it's a boundary-centric pretraining framework that scales to a 1B-parameter Vision Transformer rivaling or beating foundation models up to 7x larger — including DINOv3 — on NYU-Depth v2 RMSE and other dense spatial benchmarks. Fully open-sourced, weights on Hugging Face.

So what: Robots and physical-world agents don't need to know "that's a cat" — they need to know exactly where its edges are in 3D space to avoid it, grasp it, or navigate around it. A 1B model beating 7B-class models at that specific job, and open-sourced, is a direct unlock for anyone building robotics perception stacks without frontier-scale compute.

Hy3 is a sparse 295B MoE that activates only 21B parameters per token, plus a 3.8B multi-token-prediction layer and a 256K context window. Targets reasoning, agentic workflows, and long-context tasks. Fully open under Apache 2.0 — weights, code, all of it.

So what: A 295B-class model that only costs 21B-active-params to run puts frontier-scale reasoning within reach of teams that can't afford dense 295B inference. Combined with GLM-5.2's rise last week, China's open-weight labs are now shipping at a pace and quality that's actively reshaping where enterprises route their token spend.

Audex (Nemotron-Labs-Audex-30B-A3B) is built on Nemotron-Cascade-2-30B-A3B, a strong text-only MoE. It adds speech recognition, translation, TTS, and audio generation inside the same decoder — without regressing on the backbone's text reasoning quality, the usual tradeoff when bolting audio onto a language model.

So what: Most "unified" audio-text models pay for audio capability with degraded text intelligence. NVIDIA claims to have avoided that tax entirely, which — if it holds up under scrutiny — means one model can now replace separate ASR, TTS, and text-reasoning stacks without a quality compromise on any of them.

Antidoom uses Final Token Preference Optimization (FTPO) to target the specific failure mode where reasoning models get stuck repeating or spiraling instead of converging on an answer. Doom-loop rates dropped as far as 90% in testing — reported cases fell to roughly 1.4% — with eval scores improving across the board as a direct result of less looping. Code and blog are both public.

So what: Doom loops are one of the most expensive silent failure modes in production reasoning models — burning tokens and latency with no useful output. A lightweight, open training fix that also improves raw eval scores is a rare free lunch.

OpenScience is an Apache 2.0, model-agnostic workbench that runs the full scientific research loop — literature review, experimentation, and paper drafting — for machine learning, biology, physics, and chemistry work. Plugs into any frontier or open-weight model; runs on the researcher's own infrastructure.

So what: Every major lab has built an internal "AI scientist" tool this year, but they've stayed closed. OpenScience is the first credible open equivalent — a real threat to paid research-automation platforms, and a fast way for academic labs to get agentic research tooling without vendor lock-in.