AINov 9, 2026, 3:01 PM UTC·2 min read

Thinking Machines Lab Unveils Inkling 2 Small, Open Weights Due

Thinking Machines Lab has launched Inkling 2 Small, a 594B-parameter mixture-of-experts model with image input. Apache 2.0 weights are planned for the following week.

Abstract illustration of a large neural network with expert modules lighting up, representing the Inkling 2 Small mixture-of-experts AI model. — Image: BRIDGE

Thinking Machines Lab announced Inkling 2 Small on November 5 (US time), the second generation of its model family. It is a mixture-of-experts (MoE) model with 594 billion total parameters, of which 17 billion are active during inference. It accepts image input and supports "variable effort", which lets users adjust how much effort the model spends on reasoning.

For now, the model is available through Tinker, the company's fine-tuning and inference platform, without public weights. Once external safety testing and red-teaming are complete, the company plans to release all weights under the Apache 2.0 license the week after the announcement, along with downloadable checkpoints. Users can already try it in Tinker Playground and use it for fine-tuning and sampling on Tinker.

Architecture and performance

According to the model card, Inkling 2 Small is a 48-layer decoder-only model that routes each token to 8 of 384 experts. Images are encoded with a Vision Transformer, and the context length reaches up to one million tokens.

The model scored 31 on the Artificial Analysis Intelligence Index, well above the previous generation. The company says it improved in agentic coding, vision, knowledge work and reasoning, and is particularly strong at long-running autonomous tasks.

Combining vision and agent skills, the model can rebuild websites, games and 3D designs while checking how they look. Demonstrations include a Three.js city simulator that generates buildings and cars without ready-made assets, a Python-based PC emulator that boots a real Linux kernel, and video made in Blender.

How it was trained

The team used reinforcement learning on hundreds of thousands of long-horizon coding tasks spanning multiple agent execution environments and context-compression methods. Beyond unit-test results, it used agentic grading that evaluates how the model approaches a problem. Early checkpoints were tested on OpenCode under the name "Fledge-Alpha".

Thinking Machines Lab is an AI research and product company staffed by scientists and engineers who have worked on ChatGPT, Character.ai and PyTorch, among other projects.

Apps worth a look

SUMO: the number puzzle

Tap the right tiles, add them up, hit the target. Easy to learn, hard to master. Free, for all ages.

Dialo: learn languages by speaking

Talk to an AI conversation partner; every sentence comes back corrected and explained in your language. 9 languages, A1 to C2.

ediviewer: read EDI files like documents

Open EDIFACT and X12 messages, see every segment explained, convert to JSON/XML, export PDF and Excel. Runs locally, no upload.

giftdetect: the right gift in minutes

Pick who it is for, the occasion and their personality. The search engine reads product descriptions from local shops and ranks what fits. Free, no sign-up.

Tags:thinking machines labinkling 2 smallopen weightsmixture-of-expertstinker