Thinking Machines launched the Inkling model on July 15, 2026. The model has 975 billion total parameters and 41 billion activated parameters, supporting text, image, and audio input, with full weights available for download and fine-tuning.
Inkling adopts a mixture-of-experts Transformer architecture, with 45 trillion tokens of pretraining data covering text, images, audio, and video. It supports a context window of up to 1 million tokens, along with a preview version Inkling-Small that features 12 billion activated parameters.
The model is available via API on the Tinker platform with a context window of 256K tokens, while the open-weight version supports 1M tokens. Pricing is per million tokens: $1.87 input and $4.68 output for 64K context; $3.74 input and $9.36 output for 256K context.
Technical Implementation Path
Audio input is processed in the form of dMel spectrograms, while images are divided into 40×40 pixel patches, encoded through four layers of hMLP, and then enter a shared latent space alongside text tokens. Training covers multiple tasks such as agents, reasoning, programming, instruction following, and factual accuracy.
Inkling supports adjustable reasoning strength, allowing developers to trade off between performance and token consumption. In Terminal Bench tests, it achieves the same performance level using only one-third of the tokens compared to Nemotron 3 Ultra.
Practical Impact on Various Stakeholders
Developers can directly download the full weights from Hugging Face and deploy locally or fine-tune via the Tinker platform. Inkling Playground allows developers to interact directly with the model and evaluate its style and benchmark performance.
Enterprise users can leverage its multimodal capabilities to build applications requiring voice and visual input. On GDPval-AA v2, the model scores 1238 Elo, higher than Kimi K2.6's 1190 and DeepSeek v4 Flash max's 1189; on the τ³-Banking test, it achieves 24%, higher than Kimi K2.6's 21%.
On the Artificial Analysis Intelligence Index, Inkling scores 41, surpassing Nemotron 3 Ultra's 38, Gemma 4 31B's 29, and gpt-oss-120b's 24, making it the leading open-weight model released by a US lab. It differentiates itself in agent performance and token efficiency.
Comparable Data Comparison
Compared to GLM-5.2 max, Kimi K2.6, and DeepSeek v4 Pro max, Inkling generates an average of 25K tokens on Intelligence Index tasks, whereas the aforementioned models produce 43K, 38K, and 37K tokens, respectively.
On audio benchmarks VoiceBench, MMAU, and AudioMC, Inkling ranks among the top open-weight audio models; in visual tasks, it can call Python tools to assist with image understanding.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接