Inkling 2 Small

Product Introduction

1. Definition: Inkling 2 Small is a Mixture-of-Experts (MoE) artificial intelligence model developed by Thinking Machines Lab, featuring a massive 594 billion total parameters with 17 billion active parameters per token, coupled with a 292M-parameter vision encoder.
2. Core Value Proposition: Built for long-horizon coding, visual interpretation, agentic knowledge work, and autonomous autoresearch, Inkling 2 Small delivers frontier-level performance on complex multi-step workflows. Available on Tinker for inference, fine-tuning, and sampling, it bridges the gap between raw model intelligence and real-world execution through variable-effort reasoning and robust reinforcement learning (RL).

Main Features

1. MoE Architecture with Variable Effort: Inkling 2 Small utilizes a sparse Mixture-of-Experts architecture containing 594B total parameters while activating only 17B parameters per forward pass to maintain high computational efficiency. It supports variable-effort reasoning settings, allowing users to scale reasoning compute dynamically depending on task complexity.
2. Perception Encoder Vision Integration: The model processes high-resolution image inputs via a 292M-parameter vision encoder based on Perception Encoder architecture. Utilizing 20x20-pixel patches with 2x2 spatial pooling and interleaved text-image tokenization, it excels in OCR, visual grounding, infographic parsing, and multimodal agentic loops.
3. Long-Horizon Agentic and RL Training: Trained using Reinforcement Learning (RL) across hundreds of thousands of long-horizon coding tasks, Inkling 2 Small leverages agentic RL graders that evaluate methodology and step-by-step problem-solving rather than relying solely on traditional binary unit-test rewards.

Problems Solved

1. Pain Point: Traditional large language models struggle with long-horizon autonomous coding tasks, multi-step autoresearch, and maintaining context or verifying outputs over extended execution loops without human intervention.
2. Target Audience: Autonomous AI agent developers, machine learning engineers, software architects, data scientists, and quantitative forecasters requiring programmatic reasoning and automated ML pipeline execution.
3. Use Cases: End-to-end autonomous machine learning pipeline training (such as automated time-series forecasting and model tuning on Tinker), GPU kernel optimization (TileLang optimization), procedural web/game development via Three.js, and complex multi-modal chart/document interpretation.

Unique Advantages

1. Strengths & Limitations (Pros & Cons):
   • Pros: Exceptional performance on SWE-bench Pro (86.3% on Hard variant), state-of-the-art vision-to-code execution capabilities, advanced instruction-following, and native support for fine-tuning and checkpoint downloading via Tinker.
   • Cons: Heavy computational requirements during inference and training despite the sparse 17B active MoE design; complex autoresearch tasks still require strict supervision budgets (e.g., $500 financial or time caps) to prevent runaway execution costs.
2. Key Alternatives & Differentiation: Unlike flash models like DeepSeek-V4.1-Flash or Kimi K3, Inkling 2 Small is explicitly engineered around agentic RL graders and long-horizon autoresearch workflows rather than single-turn chat completion. Compared to closed frontier platforms like GPT-6 Astra, Thinking Machines Lab provides a structured path to open weights under an Apache 2.0 license, alongside dedicated fine-tuning support via the Tinker platform.

Frequently Asked Questions (FAQ)

1. What is the parameter size of Inkling 2 Small? Inkling 2 Small is a Mixture-of-Experts (MoE) model featuring 594 billion total parameters with 17 billion active parameters, combined with a 292M-parameter vision encoder.
2. Where can I run and fine-tune Inkling 2 Small? Inkling 2 Small is available on the Tinker platform for inference, sampling, and fine-tuning, with downloadable model weights provided under an Apache 2.0 license following external safety testing.
3. What makes Inkling 2 Small different for coding tasks? Unlike models optimized solely for static code generation, Inkling 2 Small was trained using reinforcement learning on hundreds of thousands of long-horizon coding tasks with agentic RL graders, enabling it to verify, test, and iteratively debug its work autonomously.
4. Does Inkling 2 Small support image and video inputs? Yes, it features a native multimodal vision architecture using Perception Encoder that ingests high-resolution images, charts, and spatial interfaces to power visual agents, OCR, and interactive UI generation.