On openness · The But Is It Open? blog

Scoring Kimi K3.1 Pro against the rubric

· 4 minute read

Moonshot gave away 1.56 TB of weights under a nearly-MIT license. The rubric's five questions separate what the release puts in your hands from what it lets you see — and the recipe is the half still missing.

Abstract painting on warm ivory paper: five small hand-painted ochre and burgundy cards in a loose row on the left, thin dark lines running toward an enormous deep navy painted plane on the right and stopping short of it.

Scoring Kimi K3.1 Pro against the rubric

Last week I wrote about the license that ships with Kimi K3.1 Pro — the MIT-style grant with a $20 million clause and a 100-million-user clause riding on top of it. That post asked what the terms say. This one runs the release through the book's rubric instead: five questions, Access, Permission, Deployment, Transparency, Reproducibility, and the tier spectrum from 0 to 5 that falls out of them.

First, the credit, because it is substantial. Moonshot put the complete weights of Kimi K3.1 Pro on Hugging Face on November 4: a 2.78-trillion-parameter mixture-of-experts model with a million-token context window, post-trained from Kimi K3 with large-scale joint reinforcement learning, about 1.56 terabytes of safetensors. Releases at this size are not gestures. They cost real money to train and real engineering to package, and this one was complete enough that quantized variants from third parties were appearing within the week.

Releases at this size are not gestures. They cost real money to train and real engineering to package.

Whatever the license says, the artifact itself was given away whole.

Access — a clear yes. The weights, the configuration, the inference code, and a working model card are all on the Hub, downloadable without a gate beyond a click. There are 8-bit and FP4 quantizations already, plus a Docker one-liner if you want the model without thinking about environments. Nothing about access is hedged. This is the easiest question the rubric asks and the release passes it without breaking stride.

Permission — yes for nearly everyone, with a scale tripwire. The grant is close to verbatim MIT: use, modify, merge, distribute, sublicense, sell, fine-tune. The conditions only arrive at scale — a separate agreement if you run a Model-as-a-Service business over $20 million in trailing twelve-month revenue, and an attribution requirement on products past 100 million monthly users or $20 million in monthly revenue. For a startup, a lab, or anyone hosting for customers below those lines, the license behaves like MIT. It is not OSI-approved, so I would not call it open source in the formal sense. But it grants more than most open-weight licenses do, and the thresholds are published numbers rather than vague "contact us" language. That deserves genuine credit.

Deployment — yes in principle, expensive in practice. Nothing in the license or the architecture stops you from running this model on hardware you control. The catch is arithmetic, not law: 2.78 trillion parameters means the weights alone are one and a half terabytes before you think about KV cache for that million-token context. Deployment independence exists on paper for everyone and in practice for people with serious GPUs. The quantized variants narrow the gap, and the inference providers cover the rest. I count this a pass, with the honest note that "run it yourself" is a richer person's verb at this scale than it is for a 20-billion-parameter model.

Transparency — partial, and partly labeled. The model card and the technical blog describe the architecture, the post-training pipeline, the benchmark results, and one disclosure that impressed me: Moonshot stripped cybersecurity-related training data from mid-training and reinforcement learning before the open release, and said so. What you do not get is the corpus. Data sources, mixture proportions, filtering rules — none of that is published. The rubric treats documentation of the training process as a real dimension of openness, and here the release gives you the shape of the process without its contents.

Reproducibility — no. No training code, no data lineage, no recipe a skilled group could follow to build a substantially equivalent system. This is the dimension where nearly every corporate open-weight release stops, and Kimi K3.1 Pro stops there too. Only Ai2's OLMo line and a handful of publicly funded models have gone further.

Where it lands

Tier 3 — practical open-weight. Strong on Access, Permission, and Deployment, partial on Transparency, absent on Reproducibility. It is the same tier as DeepSeek's V3 and R1 and OpenAI's gpt-oss, which is the right company: models you can genuinely take home, wrapped in terms that genuinely permit it, whose training remains something you are told about rather than shown.

The rubric's job is to keep that distinction precise. "Open" in the marketing sense covers all of these models. The five questions separate what a release puts in your hands from what it lets you see, and Kimi K3.1 Pro puts a very large amount in your hands while letting you see very little of how it was made.

So — but is it open? If openness means can I download the weights and run them where I choose, under terms that mostly leave me alone: yes, with a bill attached. If it means can I study how this thing came to be: not yet. The license turned out to be the generous half of this release. The recipe is the half still missing.

Adapted from the draft manuscript of But Is It Open?

← More from the blogJoin the newsletter ↗