Altern #20 - Kimi K3
China just put a 2.8-trillion-parameter model on the table for free, and Wall Street noticed everything else too.
Hey there — this was the week an open-weight Chinese model muscled into the frontier conversation, Google delayed Gemini for the third time, and Anthropic’s IPO math started pointing at a trillion dollars. Let’s get into it.
This Week in AI
Moonshot AI dropped Kimi K3, the biggest open-weight model ever made. 2.8 trillion parameters, a 1-million-token context window, and it beats Claude Opus 4.8 and GPT-5.5 on several coding and agentic benchmarks — trailing only Claude Fable 5 and GPT-5.6 Sol overall. Full open weights land July 27. Read more
Gemini 3.5 Pro got delayed for the third time. Google reportedly pulled it back again after the rebuilt model fell short on coding and complex reasoning, and Alphabet shares dropped about 4% on the news. One delay is normal; three in a row is starting to look structural. Read more
Anthropic’s IPO math now implies a trillion-dollar debut. Secondary markets are pricing the confidential filing at $1.05–1.15 trillion, a premium over its $965 billion private round, with an October Nasdaq target that would put it ahead of OpenAI to public markets. Read more
OpenAI gave ChatGPT unified search. As of July 14, you can search across your chats, images, documents, and projects in one place, synced across web, iOS, and Android — turning ChatGPT into something closer to a personal knowledge base than a chat window. Read more
Deep Dive
Kimi K3: the open-weight model that just closed the gap
Every few months someone claims an open model “rivals the frontier.” Kimi K3 is the first one this year where that claim actually holds up on the numbers. Moonshot AI, the Beijing startup behind the Kimi chatbot, shipped a 2.8-trillion-parameter mixture-of-experts model on July 16 — the largest open-weight model ever released, full stop.
A few things make it more than a headline:
It’s genuinely efficient. K3 only activates 16 of its 896 experts per token — about 1.8% of the total model — which is how something this large stays usable at all.
It’s cheap, for what it is. API pricing lands at $3 per million input tokens and $15 per million output, with cache hits dropping to $0.30. That’s roughly Claude Sonnet-class pricing for a model bumping shoulders with the frontier.
It’s already winning specific fights. K3 took #1 on Arena.ai’s Frontend Code leaderboard, ahead of Claude Fable 5, and beats GPT-5.5 and Opus 4.8 across several coding and agentic benchmarks. It still trails Fable 5 and GPT-5.6 Sol overall — this isn’t a “China wins” story, it’s a “the gap just got a lot smaller” story.
It’s OpenAI-SDK compatible, so switching an existing project over to it is closer to a config change than a rewrite.
How to actually try it this week:
The API is live now through Moonshot directly or aggregators like OpenRouter — useful if you want to A/B it against whatever you’re already running without juggling a separate account.
Full open weights (Modified MIT license) land July 27. If you’re set up to self-host a 2.8T MoE, that’s the date to watch.
Heads up if you’re testing it: K3 currently only supports one reasoning effort level (”max”), so it burns reasoning tokens generously — testers have seen ~13K tokens for a simple SVG generation. Budget accordingly.
Full spec breakdown if you want to go deeper: Kimi K3 complete guide.
AI of the Week
This week’s news was all about which model is smartest — so this week’s tools are about what you actually build with them.
OpenRouter — one API and one bill for hundreds of models, including new releases like Kimi K3 the same day they launch. If you want to A/B test frontier vs. open-weight without rewriting your integration each time, this is the shortcut.
Cursor — the AI-native code editor that keeps adding new frontier models (Grok 4.5 included) as soon as they ship, so you can switch models mid-project without leaving your editor.
Devin — Cognition’s autonomous software engineer, now deployed at companies like Goldman Sachs. Hand it a ticket in plain English and it plans, codes, tests, and opens the PR on its own.
Thanks for reading this week — if Kimi K3 or any of these tools end up in something you’re building, hit reply and tell us about it. See you next week.