Former OpenAI CTO Releases New Model Inkling-Small: With less than one-third of the parameters of Inkling, the performance is only 1 point lower

According to Dynamic Beating monitoring, Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, has released Inkling-Small and made the full weights open source. This is a multimodal reasoning model that supports text, image, and audio inputs. Like Inkling, the model is released under the Apache 2.0 license.

Inkling-Small has a total of 276 billion parameters with 120 billion activations per inference. In comparison, Inkling has a total of 975 billion parameters with 410 billion activations. In terms of Artificial Analysis's Intelligence Quotient, they score 40 and 41, respectively.

The smaller model actually outperforms in some tasks. In official tests, it excelled over Inkling in SWE-bench Verified, HLE, and other programming and reasoning evaluations. However, Inkling demonstrates better knowledge coverage and factual accuracy.

The price of output for Inkling-Small is $1.20 per 1 million tokens, approximately 70% cheaper than Inkling. The full weights have been uploaded to Hugging Face, but the official quantized weights still amount to 171GB, which cannot fit entirely on consumer-grade GPUs.