Thinking Machines launches Inkling 2 Small on Tinker

Thinking Machines Lab launched Inkling 2 Small on November 5, 2026, introducing a multimodal mixture-of-experts model with 594 billion total parameters and 17 billion active parameters. The company said the model is available now through its Tinker platform, but only with closed weights at launch while a broader release remains pending.

The announcement matters for founders and AI teams because Inkling 2 Small is designed not just for text generation, but for image input, visual work and agentic coding. Thinking Machines said the model can handle long-horizon tasks, and it framed the release as its second model generation. The company also said the system supports variable effort, meaning it can adjust how much computation it uses depending on the task.

What is Inkling 2 Small?

Inkling 2 Small is Thinking Machines Lab’s latest multimodal model, and the company says it accepts image inputs and generates text. It uses a 292-million-parameter vision encoder based on Perception Encoder, and Thinking Machines said it interleaves image patches with text tokens to process visual information alongside language.

The company described the model as suitable for agentic coding, visual work and long-horizon tasks. In its own examples, Thinking Machines said the system can use visual feedback to iteratively build and refine websites, games and 3D designs with interactive elements.

Thinking Machines also said Inkling 2 Small was trained with reinforcement learning across hundreds of thousands of long-horizon coding tasks using multiple agent harnesses and context-compaction strategies. The company attributed the model’s gains to more efficient pre-training and a single large unified reinforcement-learning run.

How was the model trained and benchmarked?

Thinking Machines said Inkling 2 Small’s initial pretraining phase covered 8.4 trillion tokens and ran for 68 hours on NVIDIA GB300 NVL72 racks. The company said training used NVFP4, the 4-bit floating-point format introduced with NVIDIA Blackwell GPUs.

The lab also highlighted a separate reinforcement-learning run on Tinker for time-series forecasting. In that experiment, Thinking Machines said Inkling 2 Small autonomously trained Qwen3.5-4B and more than doubled the forecasting score after 4.5 hours and $426 in Tinker spend. The company said the run had a 12-hour limit and a $500 budget, and that on unseen August and September data the final forecasting model made about 15% less error than the base model.

On its own benchmarks, Thinking Machines said Inkling 2 Small reached 86.3% on SWEBench Pro v2 (Hard), 81.6% on MMMU Pro (Standard 10), 83.8%/91.5% on ScreenSpot Pro original/with Python, and 1527 Elo on AA-Briefcase v1.1. The company also said early checkpoints were tested on OpenCode under the name Fledge-Alpha.

Artificial Analysis separately reported that Inkling 2 Small scored 31 on its Intelligence Index, up from 13 for Inkling Small. It also said the new model averaged about 150,000 output tokens per Intelligence Index task versus about 37,000 for Inkling Small, and cost about $1.18 per task versus $0.20 for its predecessor.

Why are the weights still closed?

Thinking Machines launched Inkling 2 Small with closed weights and said it would be available for inference, sampling and fine-tuning through Tinker and Tinker Playground. The company said users could try the model immediately in Tinker Playground, but downloadable checkpoints would wait until a later weights release.

Thinking Machines said it intends to release the full weights under an Apache 2.0 license after external safety testing and red-teaming. The company said that release was still pending further safety evaluations by the company and partners, and its announcement did not say the weights had already been made public.

Later reporting said the model remained proprietary as of November 6, with the weights still unreleased and access limited to Tinker. ZDNET also reported that Thinking Machines said the weights would arrive the following week, while SiliconANGLE said the company planned to release them under an open-source license the following week.

Artificial Analysis said that, if the Apache 2.0 release happens as planned, the model could become the highest-scoring open-weight model developed outside China under its rankings. That characterization remains contingent on the release actually occurring.

What happens next for developers?

For now, developers can work with Inkling 2 Small through Thinking Machines’ Tinker platform rather than download the model directly. The company said the launch includes inference, sampling and fine-tuning access, which makes the model usable for experimentation before any public checkpoint release.

That access model also gives Thinking Machines a way to gather feedback while it completes further safety reviews. The company said it still trails closed-weights models on its alignment evaluations, naming GPT-6.1 Sol as an example, which suggests the release strategy is still tied to evaluation work rather than a completed open-weight launch.

The product rollout arrives as Thinking Machines has also been the subject of outside fundraising reports. SiliconANGLE said Business Insider reported the company was raising a $5 billion round led by Accel at a $40 billion valuation, but that financing was described as a report rather than a company announcement.

For startups building on frontier models, the near-term takeaway is simple: Inkling 2 Small is live, but only in a controlled form. The weights are still pending, the company says safety work must come first, and broader distribution will depend on whether Thinking Machines follows through on the planned Apache 2.0 release.