Thinking Machines Lab frames its Nov. 5, 2026 launch under “Thinking Machines Lab: Introducing Inkling 2 Small — Nov. 5, 2026.” For related coverage, see Meta Rolls Out Muse Voice and Expressive Avatars, OpenAI Launches GPT-6 Cyber Security Model, and Meta Opens Muse Spark Weights: 816B Model Runs on Two Macs.
Key facts
| Item | Detail |
|---|---|
| Architecture (Thinking Machines Lab) | Mixture-of-experts model with 594B total parameters and 17B active per token |
| Modalities (Thinking Machines Lab) | Supports image inputs and variable compute effort |
| Artificial Analysis Intelligence Index (lab citation) | Scores 31; Thinking Machines Lab says this far surpasses prior Inkling models on that index |
| Access today (Thinking Machines Lab) | Available through Tinker while weights remain closed |
| Weights roadmap (Thinking Machines Lab) | Plans to release full weights under Apache 2.0 after external safety testing and red teaming complete |
| Tinker surface (Thinking Machines Lab) | Try in Tinker Playground; fine-tuning and sampling on Tinker; downloadable checkpoints when the weight release ships |
| Positioning (Thinking Machines Lab) | Highlights long-horizon autonomous work, agentic coding, vision workloads, knowledge work, and reasoning |
| Model card (Thinking Machines Lab) | Vision encoder about 292M parameters based on Perception Encoder; context length up to 1M tokens |
What Thinking Machines Lab describes
Thinking Machines Lab describes Inkling 2 Small as a large MoE system with multimodal inputs and a controllable effort setting for heavier jobs. It cites Artificial Analysis for a 31 score on the Artificial Analysis Intelligence Index and frames that figure within that provider’s methodology without adding separate benchmark claims in the news post.
Access runs through Tinker today, including the Tinker Playground plus fine-tuning and sampling on that platform. Weights stay closed for now; Thinking Machines Lab says it intends to publish full checkpoints under Apache 2.0 after external safety testing and red teaming finish, without committing to a day beyond that sequencing.
The announcement highlights long-running autonomous work, agentic coding, vision-heavy prompts, knowledge work, and reasoning as focus areas rather than listing every evaluation table in one release note.
What it means
Analysis (ours). Closed weights plus a stated Apache 2.0 release after safety review means on-prem rollout plans should stay tied to the actual weight drop, not the initial blog date. MoE scale and image support point at research and enterprise teams that blend documents with visuals, while Tinker supplies a managed try-before-download path.
What to watch
Progress through the lab’s safety and red-team gate, how the eventual Apache 2.0 package reads in license files, and whether Artificial Analysis revises its index as comparable MoE models ship. Tinker pilots should also watch policy changes to fine-tuning, exports, and rate limits as weights shift from closed to downloadable.