Baidu open-sources ERNIE 5.5 Flash Base: 280B weights, 14B active
The 280B MoE is downloadable, but it's a base model with a license catch. Plus a $21K open image model, a paper on agent drift, and Dropbox's Tigris deal.

Copy markdown
280B open weights, built for you to fine-tune
ERNIE-5.5-Flash-Base is a 280B-parameter MoE with ~14B active per token, pretrained on 33.5T tokens and taking context up to 1M tokens for training. Baidu shipped it as a base model — not instruction-tuned — so it's aimed at domain adaptation, and the license carries commercial-use restrictions. The weights are out; read the terms before you build on it.
A text-to-image model trained for $21K, Apache 2.0
Linum open-sourced P-JiT 21K, a 2B pixel-space diffusion transformer. The 8.7GB model.safetensors and generate.py are on Hugging Face and GitHub under Apache 2.0, and the full run cost a self-computed $21,038 (8,415 H100-hours). It's a research checkpoint — no text rendering, weak on full-body or multi-person scenes — but it's a real, downloadable base for your own fine-tunes.
Why your coding agent loses the thread
New paper Switch-SWE tested 7 coding agents on resuming interleaved, unfinished work: focus-task resolution fell an average 11.5 points (6-15 pp). A drop-in context-management module recovered 5.5-15.5 pp on OpenCode — a concrete lever if your agent runs parallel tasks.
Dropbox buys Tigris to seed an agent cloud
Dropbox says it plans to acquire AI-storage startup Tigris Data for ~$350M — its biggest deal ever — to launch 'DBX Agent Cloud,' starting with standalone object storage for AI developers. If you run on Tigris's S3-compatible storage, expect it to fold into Dropbox's agent stack.