Inkling-Small
Inkling-Small has 276 billion total and 12 billion active parameters. It accepts text, images and audio, then produces text.
What the release provides
The forty-two-layer architecture mixes local and global attention. It retains adjustable reasoning effort. Its model card describes up to one million tokens of context; Tinker serverless inference currently uses 256K.Thinking Machines: Inkling-Small model cardTinker: Current models and pricing
How to use it well
Practical guidance. Check audio accuracy with recordings that match the intended length and background noise. Compare complete task cost across effort settings instead of comparing token prices alone.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | API, Downloadable, Provider product |
| Model identifiers | thinkingmachines/Inkling-Smallthinkingmachines/Inkling-Small:peft:262144:sampling-nvfp4 |
| Context window | 262,144 tokens Tinker serverless inference beta; model-card maximum is larger Tinker: Current models and pricing |
| Input price | 0.3 USD per million tokens Tinker serverless beta, uncached input Tinker: Current models and pricing |
| Output price | 1.2 USD per million tokens Tinker serverless beta output Tinker: Current models and pricing |
| Weight license and packages | Apache 2.0; BF16 and NVFP4Thinking Machines: Inkling-Small weight repository |
| Audio guidance | 16 kHz WAV; recordings under two minutes are recommendedThinking Machines: Inkling-Small weight repository |
| API access | Serverless inference is beta; production use has a waitlistTinker: Current models and pricing |
Related models
- Inkling
Thinking Machines Lab · Inkling
A large open-weight model that reads text, images and audio.
Launch date not confirmedAPI · Downloadable · Provider product - Inkling 2 Small
Thinking Machines Lab · Inkling
A second-generation Inkling model available through Tinker.
Launch date not confirmedAPI · Provider product