Thinking Machines Lab

Sources#

What it is#

An AI research lab (publishes under "Thinking Machines Lab: Connectionism"). Mission: "build AI that extends human will and judgment." In this wiki it first appears as the org behind Interaction Models — a May 2026 research preview reframing real-time human-AI collaboration as a model-native capability rather than a harness concern. By July 2026 the strategy is legible as a three-part stack: Tinker (a hosted fine-tuning platform — anyone can customize models), interaction models (the collaboration interface), and Inkling (a from-scratch open-weights foundation family to customize).

What they've shipped / argued (as seen here)#

Interaction Models (May 2026 research preview) — models that natively take in audio/video/text and think/respond/act in real time. First model: TML-Interaction-Small (276B MoE, 12B active).
Inkling (July 2026) — their first from-scratch model, released with full weights: 975B/41B-active multimodal MoE, 1M context, controllable thinking effort, trained on GB300 systems with 30M+ RL rollouts. Explicitly not pitched as the strongest model but as the best base for fine-tuning on Tinker; Inkling-Small (276B/12B — TML-Interaction-Small's exact shape) previewed alongside. Designed to serve as the background reasoning model in the interaction-model system (Interaction / Background Model Split).
Tinker — the hosted fine-tuning platform Inkling launches on (64K/256K context, day-0 serving partnerships across Together/Fireworks/Modal/Databricks/Baseten, SGLang/vLLM/llama.cpp support). The launch demo has Inkling writing and running its own Tinker fine-tuning job. Their fine-tuned-forecaster work ("training LLMs to predict world events") feeds the Trained Calibration recipe.
Position: interactivity should scale with intelligence → it must be in the model, citing The Bitter Lesson against harness-based real-time systems (VAD, turn-detection).
Engineering footprint: upstreamed a streaming-sessions feature to SGLang; published work on defeating nondeterminism in LLM inference (batch-invariant kernels), referenced for trainer-sampler alignment; prior post On-Policy Distillation.
Running a research grant for interactivity / human-AI-collaboration benchmarks (details TBA); limited research preview of the interaction model "in the coming months," wider release "later this year"; larger models promised later in 2026.

How it connects#

Stakes out a different priority than the labs critiqued in Turn-Based Interface Bottleneck ("AI labs over-optimize for autonomy") — implicitly positioning against the autonomy-first framing seen around Anthropic's and OpenAI's agent products (Claude Code, Symphony).
Their harness-dissolves-into-model stance is the same shape as Harness Shrinkage as Models Improve (an Anthropic/Claude Code observation) — convergent thinking from different labs.
Benchmarks their model against GPT-realtime-2.0 / 1.5 (OpenAI) and Gemini-3.1-flash-live (Google) and Qwen 3.5 Omni — see Interactivity Benchmarks.

Connections#

Inkling — their first from-scratch, full-weights release (July 2026); the customization thesis made concrete
Trained Calibration — their epistemics training recipe: calibration as a first-class RL target
Interaction Models — their headline research preview
TML-Interaction-Small — the model
The Bitter Lesson — the principle they invoke
Turn-Based Interface Bottleneck — their critique of the status quo
Interactivity Benchmarks — where they benchmark against OpenAI / Google / Alibaba models
Harness Shrinkage as Models Improve — convergent thesis from Anthropic
Anthropic — peer lab; different priority ordering (autonomy-first vs. interaction-first)
Agent Harness Engineering — their interaction-models work resolves the harness-vs-model question toward the model for the real-time interaction layer

Sources#

Interaction Models: A Scalable Approach to Human-AI Collaboration
Inkling: Our Open-Weights Model — Inkling release, Tinker positioning, mission statement (vendor-claim)