Inkling 2 Small Release Notes

Thinking Machines has introduced Inkling 2 Small, a new Mixture-of-Experts (MoE) model designed for high-performance agentic coding, vision, and knowledge work. With 594B total parameters and 17B active parameters, the model demonstrates significant performance gains over its predecessor, achieving a score of 31 on the Artificial Analysis Intelligence Index.

Technical Architecture and Training

Inkling 2 Small is the first model of its scale to be trained entirely in NVFP4, a 4-bit floating-point format introduced with the NVIDIA Blackwell architecture. This choice of precision enabled rapid training cycles, with the primary pretraining phase (H1) covering 8.4 trillion tokens in just 68 hours on NVIDIA GB300 NVL72 racks.

The training process utilized a four-phase approach, culminating in a large-scale joint Reinforcement Learning (RL) run that involved 17 million rollouts over 9 days. The team employed a "StableReinforce" objective to handle staleness in their asynchronous RL infrastructure, allowing the model to scale performance log-linearly with compute.

Agentic and Vision Capabilities

Inkling 2 Small is optimized for long-horizon autonomous tasks. It integrates a 292M-parameter vision encoder based on Perception Encoder, allowing it to process image inputs for tasks ranging from OCR and infographic interpretation to complex visual reasoning.

In agentic coding, the model was trained on hundreds of thousands of long-horizon tasks using an agentic RL grader that evaluates the problem-solving process rather than just unit-test outcomes. The model shows strong performance in:

• Autoresearch: Demonstrated ability to autonomously train other models (e.g., Qwen3.5-4B) for time-series forecasting.
• Visual Reasoning: Outperformed researchers in blind wine-tasting trials using visual input alone.
• Design and Simulation: Capable of using Three.js to procedurally generate interactive 3D environments and city simulators.

Safety and Alignment

Thinking Machines has implemented a defense-in-depth safety strategy. This includes:

• Synchronous Monitoring: A system that inspects tool calls against a task-specific policy before execution.
• Asynchronous Monitoring: An LLM judge that reviews completed rollouts for misbehavior.
• Alignment Training: The use of on-policy self-distillation to internalize conduct specifications, alongside filtering of misbehavior traces during training.

While the model shows improved resistance to adversarial prompts compared to previous versions, the company notes that it still trails closed-weights models like GPT-6.1 Sol in specific alignment evaluations. The full weights are scheduled for release under an Apache 2.0 license following external safety testing.

Availability and Community Reception

Inkling 2 Small is currently available for inference and fine-tuning on the Tinker platform. Community discussion on Hacker News has highlighted the impressive velocity of the development team and the technical significance of the NVFP4 training recipe. Observers noted that the model serves as a competitive domestic alternative to other frontier open-weights models, though some users expressed concerns regarding the model's size and the naming choice.

"The most important thing here is training was entirely nvfp4 - this is best model by far that has ever been trained in that precision and explains why they were able to do it so fast." — @itkovian_

"It's a nice, domestic open weight model. Without hundreds of millions in 9 days a small team is producing a model that can beat DeepSeek 4.1 and a platform to tune it." — @epsteingpt