# Thinking Machines Lab cuts costs with Inkling-Small

_Thinking Machines Lab launches Inkling-Small, a 276B parameter model that delivers comparable performance to its larger predecessor at a fraction of the cost._

**Updated:** 2026-08-22
**Published:** 2026-07-31
**Source:** https://www.startuphub.ai/ai-news/artificial-intelligence/2026/thinking-machines-lab-cuts-costs-with-inkling-small

---

Thinking Machines Lab today announced the [Inkling-Small release](https://thinkingmachines.ai/news/inkling-small/), an efficient open-weights model designed to compete with larger systems while slashing compute requirements. This [Inkling-Small model release](/ai-news/ai-research/2026/inkling-ai-model-open-weights-multimodality) arrives as the company attempts to gain ground against competitors.

Cut CostsDriver

Thinking Machines Lab aims to gain ground against competitors by reducing compute requirementsFrom the articleThe model allows users to adjust reasoning effort, letting them choose between lower costs or higher performance.launchesInkling-Small ModelCorea 276B parameter model designed to compete with larger systems efficientlyFrom the article 7 mentionsThinking Machines Lab today announced the Inkling-Small release, an efficient open-weights model designed to compete with larger systems while slashing compute requirements.Mixture-of-ExpertsContextFrom the articleThe model functions as a mixture-of-experts transformer, utilizing 12 billion active parameters out of 276 billion total.Comparable PerformanceEffectachieves parity with the larger 975B parameter Inkling model on key benchmarksFrom the articleThe model allows users to adjust reasoning effort, letting them choose between lower costs or higher performance.StartupHub.ai ScoreContextoriginal Inkling scored 54/100, behind Intapp (70/100) but ahead of 360Learning (45/100)From the articleStartupHub.ai data assigns the original Inkling a score of 54/100.Exceeds SWEBench-VerifiedEffectFrom the articleTesting shows Inkling-Small exceeding 80 percent on SWEBench-Verified.Retains CapabilitiesEffectFrom the articleIt retains the native audio and image processing capabilities found in the larger version.Fraction of CostOutcomedelivers comparable performance to its larger predecessor at a significantly reduced costFrom the articleThe model allows users to adjust reasoning effort, letting them choose between lower costs or higher performance.

## Architecture and Efficiency

The model functions as a [mixture-of-experts transformer](/ai-news/artificial-intelligence/2026/thinking-machines-lab-interaction-models-mira-murati-2026), utilizing 12 billion active parameters out of 276 billion total. By training on NVIDIA GB300 hardware, the team claims it achieves parity with the larger 975B parameter [Thinking Machines Lab Inkling](/ai-news/technology/2026/together-ai-adds-inkling-multimodal-model) model on key benchmarks.

StartupHub.ai data assigns the original Inkling a score of 54/100. This puts it behind specialized competitors like Intapp, which holds a 70/100 rating, though it remains ahead of 360Learning at 45/100.

## Performance and Capability

Testing shows Inkling-Small exceeding 80 percent on SWEBench-Verified. It retains the native audio and image processing capabilities found in the larger version. The model allows users to adjust reasoning effort, letting them choose between lower costs or higher performance.

The company is providing full weights on Hugging Face. Developers can also access the model via the Tinker platform for fine-tuning.

---

Original analysis from [startuphub.ai](https://www.startuphub.ai), the #1 AI startup directory.