Moonshot launches Kimi K3.1 Pro, a 2.8T model with open weights due in November

What Moonshot released

Moonshot AI has launched Kimi K3.1 Pro, the first model in its K3.1 series, as MarkTechPost reported. It is an upgraded version of the company's earlier Kimi K3, trained further with much more computing power. Moonshot says it ran a single training process, called reinforcement learning (the model practices tasks and is rewarded for good results), across hundreds of different areas of work.

The model is live today on Kimi.ai, Kimi Work, Kimi Code and the Kimi API. A faster sibling, Kimi K3.1 Pro Highspeed, also shipped. According to Moonshot's blog, it runs 2 to 3x faster.

One detail matters for the headline: the model is described as open-weight, but the weights are not out yet. Open weights means the saved model files are shared so others can run and study them. Moonshot expects to release them in November 2026.

How a 2.8T model stays workable

Kimi K3.1 Pro has 2.8T total parameters. Parameters are the adjustable numbers inside a model, a bit like dials set during training. Only 104B of them are active on any one question, because the model uses a design called mixture of experts. Think of a large team of specialists where only a few are called on for each task.

In the underlying K3 design, the model picks 16 of 896 experts each time. That keeps the cost of answering far below what a model using every dial would need.

It also has a 1M-token context window. A token is a small chunk of text, and the context window is how much the model can keep in view at once. The model accepts text, images and video through the API, according to Moonshot's quickstart documentation.

Size still has a price. For K3, Moonshot recommended clusters of 64 or more specialized chips, so this is not something you run on a single graphics card.

What the scores show, and what they don't

Every number below comes from Moonshot, mostly at the model's highest reasoning setting. Competitor figures come from Kimi's own comparison chart and were not confirmed on the rivals' pages.

• Coding: 65.7% on Terminal-Bench 4.0, a test of tasks done in a command-line window. That is just behind Claude Opus 5.5 at 66.4%. On SWE-Marathon it scores 51.7, ahead of Opus 5.5 at 46.3.
• Science: The weak spot. On Terminal-Bench Science 0.1 it scores 38.1, versus 68.1 for GPT-6 Astra and 63.3 for Opus 5.5.
• Finance: 64.4 on OfficeQA Pro v2, which Moonshot calls state of the art, and 58.6 on Frontier Finance.
• Made-up answers: On an internal stress test, claims that were false fell from 23.4% for K3 to 8.3%. GPT-6 Astra still does better at 5.5%.

Several of the wins come from Moonshot's own internal tests, which outsiders cannot easily check. Independent testing will say more than this launch chart.

Price, access and the cyber question

K3.1 Pro costs $2.00 per million input tokens and $10.00 per million output tokens. Reusing text the system has already seen (a cache hit) costs $0.20. Highspeed is $4.00 and $20.00. K3 was priced at $3.00 and $15.00, so the new model is cheaper than the old one.

The interface works with the same format as OpenAI's, which makes switching easier for developers. There are some limits. Thinking mode is always on, and temperature and top-p, two settings that control how varied the answers are, are fixed. Access needs a minimum $1 top-up, and image input needs uploaded files or base64 data, not public links.

Moonshot also took cyber-related material out of training for the open-weight release. On its internal hard test of hacking-style tasks, a version without the filtering solved 90.5%, while the shipped model solves 81.0%. Moonshot says the model still trails the top proprietary models on offensive cyber work.

What to watch next

The key date is November 2026. If the weights are released as promised, outside researchers can test the claims, and companies with enough hardware can run the model themselves. If they slip, this stays a paid service from a single company.

Moonshot's own list of limits is plain: no self-hosting yet, huge serving needs, and a 25-plus-point gap on science-heavy terminal tasks. For most readers, the practical point is that the price of near-top-tier AI keeps falling.