DeepSeek V3.2, Mistral 3, and Anthropic buys Bun

What shipped

DeepSeek

DeepSeek-V3.2

Open weights with sparse attention and tool use built into the reasoning; a higher-compute Speciale variant targets hard problems.

Mistral AI

Introducing Mistral 3

Mistral Large 3, a large open mixture-of-experts model, plus small Ministral 3 models, all under Apache 2.0.

Mistral AI

Mistral Large 3

The 675B-parameter instruct weights, with about 41B active per token.

Runway

Runway Gen-4.5

A video model with sharper motion and prompt adherence that Runway positions at the top of the field.

NVIDIA

Alpamayo-R1

An open 10B vision-language-action model for autonomous driving that explains its decisions.

Anthropic

Anthropic acquires Bun

Anthropic buys the JavaScript runtime behind Claude Code and says Claude Code has reached $1B in run-rate revenue.

December opened with two big open-weight releases that landed within a day of each other.

DeepSeek V3.2

DeepSeek released V3.2 and a higher-compute variant it calls Speciale. The base model keeps the sparse attention introduced in September, and it folds tool use into the reasoning process, so it can think and call tools in the same pass. DeepSeek reports gold-level results on competition math for Speciale. I'd call that a claim to verify. What I can say without testing is that the cost-per-token story keeps improving, and every cut in price widens the set of jobs an agent can do profitably.

Mistral 3

A day later Mistral announced the Mistral 3 family: Mistral Large 3, a mixture-of-experts model with 675B total and about 41B active parameters, and three small Ministral models. Everything is Apache 2.0. Mistral is a European company that has been quieter than the Chinese labs on open weights this year, and a permissive large model from a Western lab fills a gap. Governments and regulated buyers who can't or won't use Chinese-origin weights now have a serious alternative. I'm listing both the announcement and the weights page because they point to different things: the plan and the files.

Coding economics

Anthropic acquired Bun, the JavaScript runtime that Claude Code is built on, and said Claude Code has reached $1B in run-rate revenue within about six months of general availability. I take two points from that. The agent product is already a large business. And the lab wants control of the tools under it, which is how infrastructure companies behave.

Video and driving

Runway released Gen-4.5, a video model that it says leads on its own evaluations for motion quality and prompt adherence. NVIDIA announced Alpamayo-R1, an open vision-language-action model for driving research that outputs its reasoning along with its actions, and published the weights. Models that explain what they will do next are easier to certify, a point I keep coming back to from my aviation work.

What I'd watch next

Amazon's re:Invent is this week, and the labs usually save agent announcements for conference season. A big OpenAI release before year end also feels likely.