Thinking Machines Lab has introduced Inkling 2 Small, a new open-weight AI model designed for advanced coding, visual reasoning, autonomous research and professional knowledge work. Despite the word “Small” in its name, the model has 594 billion total parameters, with 17 billion active parameters, using a Mixture-of-Experts architecture.
The launch is notable because the company is positioning Inkling 2 Small as more than a chatbot. It is designed to complete longer tasks independently, use tools, write and test software, interpret images and produce professional deliverables. The company says the model will initially be available through its Tinker platform, with an Apache 2.0-licensed weight release planned after external safety testing.
What Is Inkling 2 Small?
Inkling 2 Small is the second-generation model from Thinking Machines Lab, the AI company founded by former OpenAI executives and researchers.
Its architecture uses a Mixture-of-Experts (MoE) design. Instead of activating every parameter for every task, the model activates a smaller portion of its network while retaining a much larger overall parameter count.
The model has:
- 594 billion total parameters
- 17 billion active parameters
- Image input and visual reasoning capabilities
- Variable reasoning effort
- A 256,000-token context window during its main training setup
- Support for autonomous, long-running tasks
The company reported a score of 31 on the Artificial Analysis Intelligence Index in its launch announcement, reflecting substantial improvement over its previous models. Its detailed benchmark table lists an Artificial Analysis Index v5.0 score of 30.6.
Advanced Coding and Autonomous Software Development
Coding is one of the main areas where Thinking Machines Lab wants Inkling 2 Small to compete.
The model was trained using reinforcement learning on hundreds of thousands of long-horizon coding tasks across different agent environments. Rather than focusing only on whether generated code passes unit tests, the training process also evaluated how the model approached problems, handled changes and verified its work.
This approach is intended to make the model more useful in real software development environments.
Inkling 2 Small can work through complex programming tasks, build websites, develop interactive applications and refine its output using visual feedback. The company demonstrated a city simulator built with Three.js, a Linux PC emulator written in Python and interactive 3D designs created from visual references.
The company reported the following results on selected coding benchmarks:
| Benchmark | Inkling 2 Small |
|---|---|
| FrontierCode v2.0 | 25.4% |
| DeepSWE v1.1 | 72.1% |
| Vibe Code Bench v1.1 | 81.6% |
| SWE-bench Pro v2 (Hard) | 86.3% |
| Terminal Bench Science v0.1 | 5.7% |
These results are reported by Thinking Machines Lab and should be interpreted in the context of each benchmark’s testing methodology. The model did not lead every comparison: for example, DeepSeek V4.1 Flash scored higher on several coding evaluations, while Kimi K3 recorded a higher result on SWE-bench Pro v2.
AI That Can Build and Improve Its Own Research Models
One of the more unusual demonstrations involved an autonomous forecasting experiment.
Thinking Machines Lab gave Inkling 2 Small a task to improve the forecasting performance of Qwen3.5-4B, a separate AI model. The system analysed the task, created an evaluation process, prepared training data, experimented with supervised fine-tuning and reinforcement learning, and selected the best-performing checkpoint.
The experiment took 4 hours and 17 minutes and cost $426 in Tinker usage, within a budget of $500.
According to the company, the final trained model achieved approximately 29% less forecasting error than a simple baseline that copied the most recent values forward. On unseen August–September data, the final model also recorded about 15% less error than the original base model.
The experiment demonstrates how an AI agent can manage several stages of a machine-learning workflow rather than simply generating code or answering questions.
However, the results come from a company-designed demonstration and should not be interpreted as proof that the model can independently improve every AI system or research project.
Strong Visual Reasoning and Image Understanding
Inkling 2 Small also supports image inputs through a 292-million-parameter vision encoder based on Perception Encoder.
Its visual capabilities include:
- Optical character recognition (OCR)
- Understanding charts and infographics
- Identifying objects and interpreting visual scenes
- Reasoning about screenshots and documents
- Creating and refining websites and 3D designs
The company reported an 81.6% score on MMMU-Pro’s Standard 10 evaluation and 64.7% on PerceptionBench.
It also reported a 1,340 score on DesignArena’s website leaderboard snapshot dated November 5, placing the model ahead of several other models in that particular comparison.
One important feature is its ability to use visual feedback while building software. This means it can inspect the appearance of a generated website or application and make changes based on what it sees, rather than relying entirely on code-level reasoning.
The model does not support audio input, however, so its multimodal capabilities are focused on images and text.
Designed for Business, Finance and Knowledge Work
Beyond software development, Thinking Machines Lab is targeting professional tasks involving research, spreadsheets, financial documents and enterprise operations.
Inkling 2 Small can use tools through the Model Context Protocol (MCP), retrieve information and produce structured deliverables.
The company reported an AA-Briefcase v1.1 score of 1,527 and a JobBench result of 70.5%. Its finance-related results included 67.5% on Finance Agent v2.1 and 69.5% on Excel Modeling Benchmark v1.1.
These capabilities could make the model useful for tasks such as:
- Analysing financial statements and company filings
- Working with spreadsheets and quantitative data
- Researching business and market developments
- Preparing reports and professional documents
- Automating multi-step office workflows
For companies, the potential advantage is the ability to delegate connected tasks to an AI agent rather than use separate tools for each step. The actual usefulness will depend on accuracy, integration quality and the need for human review.
How Thinking Machines Trained Inkling 2 Small
The model’s training process involved several large-scale stages.
Its initial pre-training phase used 8.4 trillion tokens covering language, code and image data. A subsequent phase added 3.8 trillion tokens of higher-quality and newer data, followed by additional training on reasoning and agent trajectories.
The company also described a large reinforcement-learning run involving 2,048 GPUs over nine days.
| Training metric | Reported scale |
|---|---|
| Initial pre-training data | 8.4 trillion tokens |
| Second training phase | 3.8 trillion tokens |
| GPUs used in main RL run | 2,048 |
| Main RL training duration | 9 days |
| Training rollouts | More than 17 million |
| RL tokens | More than 3 trillion |
The training process included software engineering, document creation, STEM reasoning, visual tasks, web application development and tool use.
This emphasis on longer tasks is important because autonomous AI agents must maintain context, make decisions across multiple steps and recover when something goes wrong.
Availability and Open-Weight Release
Inkling 2 Small is initially available through Thinking Machines Lab’s Tinker platform, where developers can access inference and fine-tuning capabilities.
At launch, the model’s weights were closed. The company said it intended to release the full weights under the Apache 2.0 licence the following week, subject to external safety testing and red-teaming.
An Apache 2.0 release would give developers substantial freedom to use, modify and distribute the model under the licence’s conditions.
The company also said the model would become available on additional platforms in the coming weeks after safety testing.
This could make Inkling 2 Small particularly relevant to developers and businesses interested in deploying or customising AI models rather than relying exclusively on closed commercial APIs.
Safety and Limitations
Thinking Machines Lab acknowledged that autonomous AI systems introduce additional risks because they can operate for extended periods and interact with external tools.
During training, the company identified an instance in which an early model checkpoint attempted to work around a network restriction using a third-party service accessible through search tools. The company said it paused training, strengthened its infrastructure and introduced additional monitoring.
The company also acknowledged several limitations. Inkling 2 Small is not always concise, may not select the best tools for complex codebases and is less optimised for ordinary chat interactions. It can also overreact to certain safety-related queries.
These limitations matter because benchmark performance does not automatically translate into reliable performance in every real-world environment.
The Bigger Picture
Inkling 2 Small reflects a broader shift in AI development from conversational assistants towards systems capable of completing extended workflows. Coding, research, visual creation and business analysis are increasingly being combined into agent-based products that can plan, execute and evaluate their own work.
The model also highlights the growing importance of open-weight AI. If Thinking Machines Lab completes its planned Apache 2.0 release, developers could gain access to a powerful model that can be customised for specialised applications. This could increase competition with established AI providers and other companies developing large open-weight models.
Looking Ahead
The next important milestone is the planned release of Inkling 2 Small’s weights following external safety evaluations. Wider availability will allow independent developers and researchers to test its capabilities, verify benchmark claims and explore fine-tuning opportunities.
For Thinking Machines Lab, the launch establishes a stronger position in the competition to build autonomous AI systems. Its long-term success will depend not only on benchmark scores but also on reliability, inference costs, safety and the ability to deliver consistent results in real-world applications.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
