Inkling-Small

Inkling-Small lists base rates of $0.30 per million input tokens and $1.20 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 $0.30 input/$1.20 output per MTok and $0.06 cached input, with 256K hosted context. These current rates are established by the provider's machine-readable list, separate from a release-article cost comparison. Training/fine-tuning uses separately billed 64K/256K IDs; its current 50% promotion has no verified end date. Checkpoint storage is $0.10/GB/month and self-hosting has infrastructure costs.

Inkling-Small 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.

Inkling-Small is a balanced tier from Thinking Machines. It suits open-weight multimodal work, reasoning, agentic coding. The model accepts text/image/audio and generates text; video training data do not certify a video-input API. The card recommends 16 kHz WAV clips ideally under two minutes, shorter than Inkling's 20-minute recommendation. Native context is separate from Tinker's 256K hosted capacity. Continuous Cookbook effort uses [0,1); Tinker OpenAI chat supports none/minimal/low/medium/high/xhigh presets mapping to 0/.1/.2/.7/.9/.99, or floats [0,.99], default .9. OpenAI chat requires sampler checkpoint paths. Exact output cap is unverified. Serverless is beta and not recommended for intensive production use. Release benchmarks use vendor methods and effort=.99, with some coding trajectories capped at 256K; those trajectory budgets are not API output ceilings.

Inkling-Small 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.