Thinking Machines Inkling

Thinking Machines Inkling lists base rates of $1 per million input tokens and $4.05 per million output tokens. Estimate both input and output token volume and any additional charges for your workload. Tinker serverless beta lists the exact NVFP4 sampling ID at $1.00 input/$4.05 output per MTok and $0.17 cached input, with 256K hosted context. The model-specific cached rate differs from the page's generic 80%-off statement; use the exact row. Training/fine-tuning has separate 64K/256K IDs and rates, with a current 50% promotion whose end date is not published. Checkpoint storage is $0.10/GB/month; downloadable Apache weights still incur self-hosting costs.

Thinking Machines Inkling accepts about 1.05M tokens of context. That only matters if you routinely send very long documents, large codebases, or multi-turn histories that approach that limit.

Thinking Machines Inkling is a frontier tier from Thinking Machines. It suits open-weight frontier, multimodal, research. The model accepts text/image/audio and generates text; training on video does not certify video input. Native context comes from text_config.model_max_length, while Tinker serverless advertises 256K. WAV audio is 16 kHz; the card recommends clips within 20 minutes for optimal performance. Effort is continuous in the Cookbook [0,1), with presets none=0/minimal=.1/low=.2/medium=.7/high=.9/xhigh=.99; Tinker's OpenAI chat raw-float range is [0,.99], default high=.9. OpenAI chat uses sampler checkpoint paths, not the base sampling ID. No hard output cap is established. Serverless beta is not recommended by the publisher for intensive production use; vendor harness results retain their own effort/configuration.

Thinking Machines Inkling publishes open weights, but self-hosting depends on the licence, hardware footprint, quantisation quality, and serving stack. A hosted API is often cheaper until you have measured throughput and concurrency on your own hardware.