Thinking Machines Lab Releases Inkling, a 975B-Parameter Model, Leading in U.S. Open-Weight Models but Trailing China

In July 2026, Thinking Machines Lab released the Inkling model with 975 billion total parameters and 41 billion activated parameters, supporting text, image, and audio inputs and a 1 million token context. The model leads among U.S. open-weight models in agent tasks but trails top Chinese models overall.

Facts

Inkling is the first model trained by Thinking Machines Lab from scratch, using a Mixture-of-Experts Transformer architecture, with full weights open-sourced on Hugging Face. Pre-training used 45 trillion tokens covering text, images, audio, and video. The model supports fine-tuning via the Tinker platform, positioning it as a customizable foundation model. The concurrent preview version, Inkling-Small, has 276 billion parameters and 12 billion activated parameters.

Architecture Breakdown

The model adopts a 66-layer decoder-only Transformer, with MoE feedforward layers consisting of 256 routed experts and 2 shared experts, activating 6 routed experts plus all shared experts per token. The router uses sigmoid-based load balancing without auxiliary loss. The attention mechanism alternates between sliding window and global layers at a 5:1 ratio, using relative position embeddings instead of RoPE. Audio input is converted to dMel spectrograms, and images are divided into 40×40 pixel patches, encoded by a four-layer hMLP before being processed together with text tokens. Training first uses Muon to optimize large matrix weights and Adam for the remaining parameters, followed by supervised fine-tuning on synthetic data, and then over 30 million asynchronous reinforcement learning steps, resulting in a controllable reasoning intensity mechanism.

Industry Impact

For U.S. open-weight developers, Inkling provides native multimodal support and adjustable reasoning intensity, requiring only one-third the tokens of Nemotron 3 Ultra to achieve equivalent performance on Terminal Bench 2.1, reducing costs for long-chain workflows. For enterprise users, open weights allow local deployment, emphasizing data sovereignty and suiting customization needs. For Chinese model competitors, Inkling scores 41 on the Artificial Analysis Intelligence Index, higher than Nemotron 3 Ultra's 38, but still trails behind models like Kimi K2.6 overall, exposing issues with factual accuracy and a hallucination rate of 63%.

Comparison and Precedents

Similar to DeepSeek-V3's MoE design, Inkling also combines routed and shared experts, but its adjustments to attention and position encoding focus more on long-context extrapolation. During training, it referenced Kimi K2.5-generated synthetic data, indicating that Chinese and U.S. models have developed cross-dependencies in data generation. Compared to the closed-source approach, Thinking Machines Lab opted for open weights, positioning itself as a fine-tuning service on the Tinker platform rather than directly competing head-on with flagship models.

Strategic Assessment

Based on current benchmark performance and architectural details, Inkling is most likely to be adopted in enterprise customization scenarios, where developers will use the Tinker Playground to test its reasoning intensity curve to match specific latency requirements. Subsequent verification signals include score updates for new models on the Artificial Analysis platform and actual operational data from fine-tuning tasks on the Tinker platform.