Claude makes files, Qwen3-Next goes sparse, ERNIE 4.5 Thinking

What shipped

Anthropic

Claude can now create and edit files

Claude produces real Excel, Word, PowerPoint and PDF files inside a sandbox, as a preview feature.

Alibaba Qwen

Qwen3-Next 80B-A3B Instruct

80B total parameters with only about 3B active per token, using a hybrid attention design for cheap long context.

Baidu

ERNIE-4.5-21B-A3B-Thinking

A small mixture-of-experts reasoning model with tool use and a 128K context, under Apache 2.0.

Google

VaultGemma 1B

A Gemma trained from scratch with differential privacy, so it can't memorize and repeat individual training records.

MobileLLM-R1 950M

Sub-billion reasoning models aimed at math and code on small devices.

The release I tried first this week was Claude's new file creation. Anthropic added the ability to build Excel sheets, Word documents, PowerPoint decks and PDFs inside a sandbox, in preview for paid plans. Before this, an AI answer meant text I copied into a file myself.

I run a lot of the paperwork at an aviation infrastructure company. Supplier comparisons, cost schedules, a cover letter that has to match a template. The pattern is always the same: gather, calculate, format. An agent that carries it through to the .xlsx removes the last manual step. It also raises the review burden, because a polished file looks more trustworthy than it is. I keep a human check on every number that leaves the building.

Qwen3-Next and the sparse trend

Alibaba's Qwen team released Qwen3-Next-80B-A3B. It has 80B total parameters but activates only around 3B for each token, and it mixes linear attention with standard attention so long inputs cost less. The claim is much higher throughput on long context at similar quality to their bigger dense model. I can't verify the speed numbers myself yet, but the direction is clear: the labs are racing to make long-context inference cheap, and open weights are keeping pace.

Baidu followed with ERNIE-4.5-21B-A3B-Thinking, a compact reasoning model under Apache 2.0 with tool calling and a 128K window. Small active-parameter counts are what make self-hosting realistic on one GPU.

Privacy and phones

A quick note on why I group these together. Each of them is small, cheap and runs close to the user, and that is where I expect the next wave of useful business agents to live.

Google released VaultGemma, a 1B model trained with differential privacy from the start. For regulated work that is the interesting part. A model that provably limits memorization is easier to defend to an auditor, and auditors were my first career.

Meta added MobileLLM-R1, reasoning models under a billion parameters. They won't replace a frontier model, but they point at agents that run on the device in your hand.

What I'd watch next

OpenAI's coding agent. There has been talk of a Codex update built on GPT-5, and I want to see whether it can stay on a task for hours.