Axion 1.0, Antigravity chemistry skills, StateBench, Palette

Four primary pages I could open and read today: an open agent model, a skills pack for a coding agent, a benchmark, and a small model that builds slide decks. All numbers below come from those pages, not from me.

Axion 1.0-35B, open weights from ByteDance

The Hugging Face card describes a reproduction checkpoint of Axion 1.0, released under Apache 2.0 with a paper, code and training data alongside. It is a mixture-of-experts model with roughly 35B parameters in total and 3B active, built for scientific reasoning, information seeking and tool use.

The recipe is the interesting part. About 18K SFT examples go through two stages, one on self-contained scientific problems and one on verifiable tasks that need outside information and tools, followed by reinforcement learning over agent trajectories. There is no mid-training, DPO or on-policy distillation. The reproduction scores 52.8 on HLE, 35.3 on BMB, 60.4 on SEAL-0, 45.9 on HiPhO, 85.0 on FrontierScience Olympiad and 55.0 on FrontierScience Research. The card serves it with vLLM on eight GPUs, so it is open but not small to run.

Cheminformatics Skills for Antigravity

Google DeepMind published a suite of skills for Antigravity grounded in RDKit, with connectors to ChEMBL and PubChem. The split is clear: the language model orchestrates and reasons, while deterministic chemistry tools do the molecular calculations. That matters because changing one atom describes a different compound.

The worked example starts from crizotinib and lorlatinib, a known case where cyclizing a floppy molecule improved potency. DeepMind says the agent retrieves thousands of compounds from the literature, filters out buffers and solvents, runs matched molecular pair analysis and proposes novel rigidified macrocycles with synthetic routes in about an hour, against multiple days per iteration by hand.

StateBench from Ai2

StateBench grades state tracking at four levels, from fixed values up to variable swaps, and ties them to Krohn-Rhodes theory. Gemini 3.8 Flash, Claude Opus 5.5 and GPT-6 Sol and Luna all do well on the two easy levels, then drop to 13% or lower on the longest general finite problems.

Models trained from scratch on the task tell a different story. Gated DeltaNet and hybrid models trained on up to 256 assignments held up to tens of thousands, while transformers never learned the hardest level. Ai2 reads that as a training gap more than an architecture one for the recurrent designs.

Palette from IBM Research

IBM fine-tuned a 20B open-weight model so a designer step plans the deck and a coder step draws each slide, running the coder calls in parallel for a 6 to 7x speedup. Its fix for invented facts was reference decks in the teacher prompt, not more rules.

What I'd watch next

Whether anyone reproduces the Axion table independently, and whether StateBench numbers appear on the next batch of hybrid releases.