The news, 365 days behind — on purpose Delayed live · replaying 2025

One Year Ago.AI

Remember how fast this is.

22JUL2025replayed
one year on
open weightsAlibaba Qwen

Alibaba drops Qwen3-Coder, its most agentic open model

A 480B-parameter mixture-of-experts coding model (35B active) with claimed parity to Claude Sonnet 4 on agentic tasks.

Alibaba’s Qwen team releases Qwen3-Coder-480B-A35B-Instruct — a mixture-of-experts model with 480 billion total parameters, 35 billion active — built specifically for agentic coding: long-horizon, multi-turn tool use across repositories, with a 256K context window extensible to a million tokens.

The claimed results put it at the top of open-model coding benchmarks and in the neighborhood of Claude Sonnet 4 on agentic evaluations — the comparison every coding model now measures itself against. It ships alongside an open-sourced CLI tool, an unsubtle signal about where Qwen thinks coding is going: not autocomplete, but delegation.

One year later — open only if you can handle spoilers

Over the following year, open-weights models from Chinese labs would cycle through the leaderboard at breakneck speed, making the notion of a single 'default' open model obsolete.

Replay thisPost on XRedditHNLinkedIn

The Weekly Replay · free by email

This week, one year ago — every Sunday.

One email each Sunday: the week's replayed AI news, with the one-year-later annotations included. Written like it's breaking — dated like it isn't.

Free · double opt-in · unsubscribe anytime · privacy