Inkling: 975 B‑Parameter Open‑Weight Model from Thinking Machines Lab Targeting Customizable AI

Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, has launched Inkling, its first self‑developed large‑scale AI model with fully open weights, enabling developers and enterprises to download and customize the model directly.

Model specifications

Inkling uses a hybrid‑expert Transformer architecture, totaling 975 billion parameters with 410 billion active parameters, and supports a context window of up to 1 million tokens.

Training data and process

The model was pretrained on 45 trillion tokens that span text, image, audio, and video modalities. Training employed a mixed‑optimizer strategy: large matrix weights were optimized with the Muon optimizer, while other parameters used Adam. Hyper‑parameter scheduling borrowed from the team’s prior modular‑manifold research, and weight decay was tied to the square of the learning rate, which helped maintain stable weight scale across training stages.

Post‑training covered domains such as mathematics, agent programming, tool use, audio, vision, dialogue, and safety. An initial supervised fine‑tuning round used synthetic data generated by open‑weight models (including Kimi K2.5), after which the majority of compute was devoted to large‑scale asynchronous reinforcement learning on both synthetic and human‑crafted environments.

Architecture details

The MoE design follows DeepSeek‑V3, featuring 256 routing experts and 2 shared experts per MoE layer; each token activates 6 routing experts. Routing uses a sigmoid mechanism with a load‑balancing bias and joint normalization of routing and shared expert scores. Attention layers alternate in a 5∶1 ratio between sliding‑window and global attention, each with 8 KV heads. Relative position embeddings outperform RoPE for longer sequences, and short convolutions are inserted after the Key‑Value projection and before the MLP residual branch.

Inference intensity and benchmark performance

Inkling allows developers to adjust inference intensity, balancing performance and token efficiency. In Terminal Bench 2.1, Inkling achieves the same performance as Nemotron 3 Ultra while using only one‑third of the tokens. Similar efficiency is shown across other benchmarks such as HLE, IFBench, VoiceBench, MMAU, AudioMC, and FORTRESS, where Inkling ranks among the strongest open‑weight models for audio and safety.

Inkling‑Small preview

A lighter variant, Inkling‑Small, contains 276 billion total parameters with 12 billion active parameters. Despite its reduced size, it matches or exceeds Inkling on many benchmarks, making it suitable for latency‑sensitive workloads such as programming assistance, evaluation scoring, and synthetic data generation.

Customization via Tinker

Inkling is available on the Tinker platform with context lengths of 64 k and 256 k tokens. The Inkling Playground lets users interact with the model directly. The Tinker Cookbook now includes three example pipelines showcasing Inkling’s audio capabilities, and the tml‑renderer tool improves sampling reliability for multimodal and tool‑calling scenarios.

Inkling can be accessed through APIs from Together AI, Fireworks, Modal, Databricks, and Baseten. The team also collaborates with RadixArk (SGLang, Miles), Inferact (vLLM), Lightseek (TokenSpeed), Unsloth (llama.cpp), and Hugging Face (Transformers integration).

Model release

The full model weights, including the original checkpoint and an NVFP4 checkpoint optimized for NVIDIA Blackwell systems, are published on Hugging Face.