Your model doesn't learn, it just reads its notes very fast -- I'd like to change that.

I train open-weight models, and build the agent systems that run them. The rest of my time goes to the other half — models that keep learning after training stops.

Merged

• transformers #41420 — a routing-weight dtype cast that crashed LoRA fine-tuning of Qwen3-VL-MoE, and the same cast in five further MoE implementations
• transformers #45124 — the missing tensor-parallel plan for Qwen3.5-MoE, so tp_plan="auto" shards the head the way the text model already did
• openai-agents-python #2196 — session settings and a history limit across the SDK's six session backends

Open

• titans-pytorch #64 and #65 — per-head memory parameters share one storage through a stride-0 view. Diagnosis and fix for a crash that has been open since February 2025

atlas-torch — an implementation of ATLAS (Behrouz et al. 2025) built on lucidrains' titans-pytorch, with a five-arm controlled campaign at 170M parameters. The result is a null and the README says so: at that scale the memory helps local prediction and nothing recoverable survives past the attention window.