For Thinking Machines, from 🤫 hussh
Two halves of the same sentence.
Work backwards from the human. Thinking Machines builds the science of customizable, reliable, human-shaped intelligence. 🤫 builds the place it should live. Here is how the two roadmaps meet.
What we build together
Their theme, what 🤫 builds, and what it unlocks.
🤫 Puppy One
Bring intelligence to where the knowledge lives. AI must be distributed to benefit from distributed knowledge.
A personal supercomputer the person owns, at home or in the garage, so intelligence lives where their data already is.
Distributed AI with a physical address: millions of homes.
🤫 Private Agent One
Human participation is a technical challenge. The interface must invite and reward it.
A private agent whose interface is consent. It shows its work, asks, and keeps a receipt for every access.
People who participate because they stay in control.
🤫 Factory One
Decentralized alignment: values belong across many people and models, not a handful of places.
An edge-supercomputing grid owned by ordinary people. Decentralized ownership, not only decentralized models.
Alignment made durable by who owns the machines.
🤫 Private Agent One + Puppy One
Efficient fine-tuning so anyone can teach a model their own expertise (Tinker, LoRA Without Regret).
Teach-your-own private model on compute the person owns. Their expertise stays theirs, by consent.
Tinker's science reaching the people who own the hardware it runs on.
🤫 Agent One + Tag One
Real-time, multimodal interaction between people and machines.
An always-on, voice-first agent, and a presence wearable for the people you love.
Interaction models with a daily, intimate home.
The bench
The frontier models, through open standards.
🤫 Agent One is provider-neutral by design, Google and Gemini first on day zero, with every frontier model reachable over the open Model Context Protocol. So a person's agent is never locked to one vendor, and teaching your own model runs on compute you own. Versions as of July 2026.
Google: Gemini
Gemini 3.1 Pro and Gemini 3.5 Flash, plus the open-source Gemini CLI. Our day-zero default.
Thinking Machines: Tinker
The fine-tuning API for teaching open-weight models your own expertise. The science we most want to run on owned compute.
Anthropic: Claude and Claude Code
Claude Opus 4.8, Sonnet 5 and Haiku 4.5; Claude Code and the Agent SDK; a first-class MCP host.
OpenAI: GPT-5.6 and Codex
GPT-5.6 in its Sol, Terra and Luna tiers, and the Codex agentic coding system.
xAI: Grok 4.5
grok-4.5, xAI's flagship on the xAI API.
The open standard: MCP
Model Context Protocol (spec 2025-11-25), donated to the Linux Foundation's Agentic AI Foundation and supported across the major labs. It is what makes provider-neutral real.
Everyone wins
Good for the person, the lab and the world.
The person
Their expertise becomes their own private model, on their own supercomputer. Leverage, with their data staying home.
Thinking Machines
Their science of customizable, reliable models reaches the people who own the compute it runs on. Distribution true to their own thesis.
The world
Intelligence owned and distributed, not concentrated. Decentralized alignment made durable by decentralized ownership.
The people
The people building Thinking Machines.
From their own pages and public profiles.
Mira Murati
Founder & CEO
Former OpenAI CTO, building an independent, research-first lab devoted to AI that extends human will and judgment.
John Schulman
Co-founder & Chief Scientist
OpenAI co-founder and an RLHF pioneer. His post-training work is the heart of teach-your-own-model.
Soumith Chintala
Chief Technology Officer
Co-creator of PyTorch and one of open source's most trusted voices.
Horace He
Research, inference and GPU kernels
Deterministic inference and kernel work, the foundation of reliable inference across a distributed fleet.
Kevin Lu
Research, RL and post-training
Efficient post-training and distillation: low-cost model customization on owned hardware.
Jeremy Bernstein
Research, optimization theory
Training dynamics and optimization, the efficiency research behind a low cost per watt.
Distributed knowledge, distributed machines