Thinking Machines releases Inkling as the top-performing US open weights AI model
Thinking Machines has launched Inkling, a multimodal model that currently ranks as the highest-performing open weights AI developed by a U.S.-based laboratory.
1. Model Release and Specifications
Thinking Machines has launched Inkling, its first production-grade language model. The model features 975 billion total parameters with 41 billion active parameters. It is designed to process text, image, and audio inputs, utilizing a hierarchical patch encoder for visual data and discrete token encoding for audio. The model is available via the company’s Tinker platform API with a 256K context window, while the open weights version, accessible on HuggingFace, supports a 1M context window.
2. Performance and Benchmarks
Inkling debuted at 41 on the Artificial Analysis Intelligence Index, positioning it as the leading open weights model from a U.S.-based laboratory. It outperformed previous benchmarks, including Nemotron 3 Ultra, Gemma 4 31B, and gpt-oss-120b. The model demonstrated notable strength in agentic performance, achieving an Elo of 1238 on the GDPval-AA v2 benchmark and scoring 24% on the 𝜏³-Banking test, surpassing results from Kimi K2.6 and DeepSeek v4 Flash. Additionally, Inkling shows high token efficiency, averaging 25K output tokens per task compared to significantly higher averages from competing models.
3. Pricing and Availability
Access to the model is provided through the Tinker platform with tiered pricing based on context window size. For a 64K context window, the cost is $1.87 per million tokens for input, $0.374 for cached tokens, and $4.68 for output. For a 256K context window, the pricing increases to $3.74 per million tokens for input, $0.748 for cached tokens, and $9.36 for output. The model weights are currently available for download on HuggingFace.