TEST RECORD · ← all field reports

FIELD REPORTS/VIDEO/ISSUE #92

Reasons through your code, documents, images, and video at once

The gap between open models and the very best paid ones has never been this thin. Moonshot's new flagship is the clearest proof yet.

MODELKimi-K3
PUBLISHEDAugust 17, 2026
READ TIME3 min
TESTED BYNeural Expedition
CATEGORYVIDEO

Field notes

01What it does

You give it everything at once: the codebase, the meeting notes, the screenshots, even an hour of video, and ask your question. It reads and watches all of it in one conversation, thinks the problem through step by step, and comes back with the answer or the finished work: a patched repository, a research write-up with charts, a worked-out plan.

Two things make that possible. Its memory is vast: it holds around a million tokens in one session, roughly a dozen novels' worth, so a whole project fits without chopping it into pieces. And it does not stop at reading. It plans, uses tools, runs long stretches of work on its own, and checks its own results as it goes.

This is Moonshot's flagship and the largest open-weight release to date: a 2.8 trillion parameter system in which only a small fraction wakes up for each word. The weights are public under Moonshot's own Kimi K3 license rather than a standard open license, so companies should read the terms before building on it.

02How to try it

The fastest way in is the Kimi chat at kimi.com, where this model answers for free. Give it a real workload, not a toy: paste a long report together with the screenshots that go with it, and ask one question whose answer lives across both. Watch whether it keeps every detail straight in a single pass; that combination of memory and reasoning is the whole act. If you build things, the weights are on Hugging Face and the model is served through an API by the usual hosted providers.

03Caveat

Do not plan on running this one locally: at 2.8 trillion parameters it needs server-room hardware, so open weights here means you can study, host, or fine-tune it on serious infrastructure, not load it on a gaming GPU. For most readers the chat and the API are the product.

04What you can do with it

  • Ask one question that spans a contract, a spreadsheet export, and the screenshots of the email thread around them.
  • Point it at a large codebase and let it carry a real change through, including running the tests.
  • Turn a folder of research material into a written report with charts.
  • Ask what happens in an hour-long recording without watching it again.
  • Study how a top-tier model is put together; the full weights and technical report are public.

Try the demo

View model page

Read this issue on neuralexpedition.com →