Journal24 October 20264 min

Open weights are close. Read the licence first.

Three large open-weight models arrived in October, a few points from the closed frontier. What decides whether one can become your model is not its score. It is the licence file.

Our Formula is simple: an open base model, plus the work your people approve, equals a model that belongs to you. The first term matters more every month, because the open base models keep getting better. October showed how much, and it also showed where the catch is.

What arrived

Meta, Muse Spark Open, 15 October: 816 billion parameters, 42 billion active per token, a context of a million tokens. Self-hosting needs at least 700 GB of GPU memory at 4-bit precision.
Reflection AI, Beam, 16 October: 501 billion parameters, 23 billion active, under Apache 2.0, with Dell and HPE offering it inside a company’s own walls.
Moonshot, Kimi K3.1 Pro, 23 October: in its apps and API now, with full weights promised by 4 November. On one public index it sits three points below the best closed model.
Mistral Large 4, preview on 6 October: an API first, with weights promised by the end of the month and no licence named yet.

Close is not the same as usable

A base model is only a base for you if three things are true: the weights are in your hands, the licence lets you change them and keep the result, and the model does well on your work rather than on a leaderboard.

On licences, October was uneven. Beam’s Apache 2.0 has no revenue trigger. Muse Spark Open allows internal use and use inside your own product, but a group with more than USD 20 million in yearly revenue must notify Meta and sign a separate agreement before offering the model as a service, and products must say they are built with it. Whether a provider running a client’s own fine-tuned version counts as a service is a question for lawyers, before any training starts.

Two of the four are open in the future tense: weights promised, licence to follow. You cannot plan a model you will own on a promise.

How we choose a base

The licence file comes first. Published, read, and checked for what it allows with a fine-tuned derivative, before any test.
The weights are downloaded and pinned. We record the exact file and its fingerprint, so the base under your model never moves.
Your test decides, not the index. Candidate bases are scored on a sealed set from your own work, next to a simple baseline.
Smaller often wins. A base that fits the machine you already have, and is good at your task, beats a larger one you have to rent.

The Model Deed records the base model and its licence on the first line, for that reason: ownership starts with what you built on. The Deed is in development; the rule is how we work today.