Inkling 2 Small

Inkling 2 Small is Thinking Machines Lab’s second-generation flagship model, announced on 2026-11-05. Official post: https://thinkingmachines.ai/news/inkling-2-small/ .

Positioning: a Mixture-of-Experts (MoE) model aimed at agentic coding, vision, knowledge work, and reasoning, with image inputs and variable effort.

Highlights:

• Scale: 594B total / 17B active parameters;
• Third-party index: Artificial Analysis Intelligence Index about 31 (AA Index v5.0 reported as 30.6 in the lab’s tables);
• Broad gains over the first generation, especially on long-horizon autonomous work, vision-in-the-loop creation, and agentic coding;
• Available now on Tinker for inference and fine-tuning with closed weights; full weights planned under Apache 2.0 after external safety testing and red-teaming;
• Try it today in the Tinker Playground.

Problem

Open-weight teams need large capacity without activating every parameter, plus models that fine-tune well for real agent loops. Inkling 2 Small keeps a large total parameter count with a smaller active budget and trains heavily on long-horizon agentic tasks.

Training & design notes

• Multi-phase pretraining (including ~8.4T tokens in the first phase) run natively in NVFP4 on NVIDIA GB300 racks;
• One joint RL run: 17M+ rollouts over ~9 days on 2,048 GPUs, covering coding, office/knowledge work, STEM, web/app building, vision, and tool use;
• Vision via a ~292M Perception Encoder–style encoder with interleaved image patches and text;
• Tuned for customizability: the lab reports faster learning and higher final performance than prior Inkling models under matched fine-tuning setups.

Limits

• Starts closed-weight; Apache 2.0 release waits on safety checks;
• No audio support; chat UX is not the primary optimization target;
• Alignment metrics still trail leading closed-weight models on some internal/external suites.

Who it's for

• Teams that need inference, sampling, or domain fine-tuning on Tinker;
• Developers and labs tracking open-weight agentic coding, visual creation, and knowledge-work models;
• Groups that want downloadable checkpoints once weights ship.

Getting started

1. Read https://thinkingmachines.ai/news/inkling-2-small/ ;
2. Try inference in the Tinker Playground; use Tinker for fine-tune / sampling;
3. Watch for the post–safety-test Apache 2.0 weight drop and downloadable checkpoints;
4. When comparing numbers, note official evals often use effort 0.99 (xhigh), temperature 1.0, and up to a ~1M-token trajectory limit.