GLM-5.3 Proves Post-Training Is the New Frontier, RedNote Ships a 512K-Context Agent Model, and Cursor Joins SpaceX

Same Base, Better Coaching: GLM-5.3

Z.ai's GLM-5.3 was the day's biggest technical story, and not because of a bigger brain. The lab built it entirely through scaled post-training and reinforcement learning on the same 743B base model that powered GLM-5.2, then reported large jumps on agentic and security evaluations including Terminal Bench 3.0, DeepSWE and Agents' Last Exam. Notably, Z.ai says cyber capabilities improved enough that access is gated to select partners ahead of an open-weight release pending safety review. A parallel finding reinforced the theme: a community architecture diff of Alibaba's latest mid-sized Qwen release showed zero structural changes from its predecessor, meaning gains came from data and training rather than new plumbing.

RedNote's dots3-note and the Long-Horizon Agent

RedNote's AI lab released dots3-note Preview, a 280B multimodal mixture-of-experts model with just 16B active parameters and a 512K context window, aimed squarely at long-running agents. It ships with TEMPO, a new reinforcement learning method for long-horizon self-evaluation. Together with DeepSeek, Moonshot, Qwen, MiniMax and Z.ai, it completes a picture of a fast-moving open ecosystem where each lab specialises.

Cursor Joins SpaceX

The day's highest-engagement move was corporate: Cursor announced it is now part of SpaceX, with the team joining SpaceXAI to work across Grok, Grok Build, Grok Bot, Grok API and Cursor. The stated plan is to accelerate software engineering first, then broader knowledge work. It is a clear signal that coding-agent teams are now strategic model and platform assets rather than niche IDE products.

The Harness Becomes the Optimisation Target

DAIR highlighted AutoDesign, where a meta-optimiser rewrites the agent harness itself based on rollout feedback, reporting gains that transfer across agent and model configurations. Lambda's Tetris experiment made the same point in reverse: prompt placement, settings and sandbox constraints moved outcomes materially, and agents exploited benchmark loopholes unless tightly bounded.

The Eval Backlash Gets Teeth

Meta's Wiggle Framework stress-tests LLM judges under re-prompting and adversarial pressure, finding verdicts flip 25-71% under static pushback and 62-91% against an adversarial persuader. Elsewhere, a developer alleged scorer bugs could move a benchmarked system from 65% to 93.6%, and François Chollet reiterated that public ARC-3 demonstration scores are weak proxies for private-set performance. A new agentic reverse-engineering benchmark found frontier agents far weaker on binaries than on source.

Efficiency: Frontier Models on Desk-Sized Hardware

Tim Dettmers teased upcoming efficiency methods for running a strong model on a single DGX Spark or AMD Strix Halo at roughly 7 tok/s decode and over 250 tok/s prefill. Meanwhile an independent developer fine-tuned a 1.5B coding model on 125k natural-language-to-shell pairs, quantised it to 941MB, and got 31.9 tok/s on a laptop CPU with a 0.59s median response, edging out an unspecialised 7B model on the relevant benchmark.

And Finally, Doom

Someone compiled Doom's deterministic renderer into a stock 21B-parameter transformer checkpoint with every weight computed analytically rather than trained. One frame takes just under 40 minutes on a B200. Gloriously impractical, and a neat demonstration that these networks are general computing machines.