The mannequin features a reasoning-effort setting that builders can regulate from 0.2 to 0.99. Pondering Machines stated the setting permits customers to stability efficiency towards the variety of generated tokens. Within the firm’s testing, Inkling matched Nemotron 3 Extremely’s Terminal Bench 2.1 rating whereas producing about one-third as many tokens.
Builders can fine-tune Inkling by means of Tinker utilizing context lengths of 64,000 or 256,000 tokens and check it by means of the Inkling Playground. The mannequin is on the market by means of APIs from Collectively AI, Fireworks, Modal, Databricks, and Baseten. It’s also supported by inference software program, together with SGLang, vLLM, TokenSpeed, llama.cpp, and Hugging Face Transformers.
Inkling’s full weights can be found on Hugging Face as the unique checkpoint and as a quantized NVFP4 checkpoint. Pondering Machines additionally previewed Inkling-Small, which has 276 billion whole parameters and 12 billion lively parameters. The corporate stated it might launch the smaller mannequin’s full weights after finishing testing.