# Kimi K2 Thinking, Google's space-compute idea, and Nested Learning

> Week of 3 Nov 2025: Kimi K2 Thinking, Project Suncatcher, Google's Nested Learning, and the Maya1 open voice model.

Published 13 Oct 2026. Coverage: 3 Nov 2025 to 9 Nov 2025. By Ali Asif.

## Items

1. **Kimi K2 Thinking** (Moonshot AI, Open weights): An open reasoning model that Moonshot says can chain 200 to 300 tool calls in a row without drifting. [Primary source](https://huggingface.co/moonshotai/Kimi-K2-Thinking)
2. **Towards a future space-based, scalable AI infrastructure** (Google Research, Science): Project Suncatcher studies solar-powered satellites carrying TPUs and linked by optical connections. [Primary source](https://research.google/blog/exploring-a-space-based-scalable-ai-infrastructure-system-design/)
3. **Introducing Nested Learning** (Google Research, Science): Proposes treating a model as nested optimization problems, aimed at continual learning without forgetting. [Primary source](https://research.google/blog/introducing-nested-learning-a-new-ml-paradigm-for-continual-learning/)
4. **Maya1** (Maya Research, Open weights): A 3B open text-to-speech model where you describe the voice you want in plain words. [Primary source](https://huggingface.co/maya-research/maya1)

Four items this week, which is fewer than usual, and I'd rather give you four I'm sure about.

## Kimi K2 Thinking

Moonshot released K2 Thinking, the reasoning version of its trillion-parameter model, as open weights. The team says it can execute 200 to 300 sequential tool calls without human steering and reports strong results on agentic search and coding benchmarks. Those are the company's numbers, and I'd treat them as a starting point.

What I do care about is the shape of the capability. Most useful business work is not one clever answer. It is a long chain: read the tender, check the spec against the standard, pull three supplier quotes, flag the gaps, draft the questions. A model that holds a plan across hundreds of steps without losing the thread is the difference between a demo and an employee. Moonshot released it with open weights, so I can imagine running it inside a company's own walls for work that can't leave the building. Cost and speed at that size is the real question, and I haven't measured either.

## Compute in orbit

Google Research described Project Suncatcher, a research study on putting TPUs on solar-powered satellites linked by free-space optics. It's early, with prototypes planned in 2027, and I won't pretend to evaluate the physics. I'd note why a lab is writing about it at all. Power is the limiting input for AI, and people with a lot of sun know that better than most. I spend my days around airfield lighting in a region with plenty of daylight, so the logic is easy to follow.

## Learning that doesn't forget

Google Research also introduced Nested Learning, a way of viewing a model's architecture and its training as layered optimization problems, with a prototype called Hope. The target is continual learning, where a model absorbs new information without erasing what it knew. Anyone who has watched an agent forget a rule from three steps ago understands why that matters. It's a paper, so adoption is a long way off.

## Voices

Maya Research released Maya1, a 3B text-to-speech model that takes a plain-language description of the voice, such as age, tone and accent, and open weights. For Arabic and English voice work the open voice lane is worth watching, since commercial voice APIs are the most expensive part of video production.

## What I'd watch next

Google has been hinting at a big Gemini release. If it arrives, I'd look first at coding and long-horizon tool use.

## My take

Kimi K2 Thinking is the one that changes my planning. A long chain of tool calls from a model I can host is what makes a self-contained agent practical, and it comes from outside the three US labs. I'd compare it with Claude on a real procurement task before I make any claim about which is better.

## Sources

- https://huggingface.co/moonshotai/Kimi-K2-Thinking
- https://research.google/blog/exploring-a-space-based-scalable-ai-infrastructure-system-design/
- https://research.google/blog/introducing-nested-learning-a-new-ml-paradigm-for-continual-learning/
- https://huggingface.co/maya-research/maya1

Canonical: https://heyaliasif.com/ai/2025-11-03-week-in-ai