Thinking Machines Releases Inkling Open Weights Model
- •Thinking Machines released Inkling, a 975B parameter multimodal model, on July 15, 2026.
- •Inkling debuted at 41 on the Artificial Analysis Intelligence Index, surpassing all U.S. open weights peers.
- •The model demonstrates high agentic performance and token efficiency compared to existing open weights leaders.
Thinking Machines released Inkling, a new multimodal model, on July 15, 2026. This marks the company's first production language model release, featuring 975B total parameters and 41B active parameters. The model supports text, image, and audio inputs, with weights available on HuggingFace featuring a 1M context window, while the Tinker platform API offers a 256K context window.
Inkling debuted at 41 on the Artificial Analysis Intelligence Index, outperforming other U.S. open weights models including Nemotron 3 Ultra (38), Gemma 4 31B (29), and gpt-oss-120b (24). The model shows high agentic performance, achieving an Elo of 1238 on GDPval-AA v2, surpassing Kimi K2.6 (1190) and DeepSeek v4 Flash (1189). Additionally, it recorded 24% on the 𝜏³-Banking benchmark, compared to 21% for Kimi K2.6 and 23% for DeepSeek v4 Flash.
In terms of efficiency, Inkling averages 25K output tokens per Intelligence Index task, notably lower than the 43K, 38K, and 37K averages for GLM-5.2, Kimi K2.6, and DeepSeek v4 Pro. Multimodal capabilities include hierarchical patch encoding for images and video, plus discrete token encoding for audio. Pricing for the 256K context window is set at $3.74 per 1M input tokens and $9.36 per 1M output tokens. Inkling further scored 40% on accuracy and 63% on the hallucination rate, with a +2 score on AA-Omniscience.