TL;DR
Thinking Machines Lab released Inkling-Small, a 276B-parameter open-weight mixture-of-experts model with 12B active parameters that outperforms its larger 975B sibling on agentic coding tasks. The model achieves 80.2% on SWE-Bench Verified, 89.5% on GPQA Diamond, and runs on consumer hardware with its efficient architecture.
The Model
Inkling-Small represents a new approach to open-weight models:
- Total parameters: 276 billion
- Active parameters: 12 billion (Mixture of Experts)
- Architecture: Multimodal MoE with reasoning capabilities
- Context window: Up to 1 million tokens
- License: Open weights (Apache 2.0 equivalent)
- Modalities: Text, images, audio
The model is designed for efficiency without sacrificing capability. By activating only 12 billion parameters per inference, Inkling-Small can run on consumer hardware while delivering performance that rivals much larger models.
Benchmark Results
Independent testing showed strong results:
- SWE-Bench Verified: 80.2% (outperforming its 975B sibling)
- GPQA Diamond: 89.5%
- IFBench: 82.2%
- Humanity’s Last Exam (text-only): 31.6%
The coding performance is particularly notable. At 80.2% on SWE-Bench Verified, Inkling-Small matches or exceeds many closed frontier models on real-world software engineering tasks.
Why It Matters
Inkling-Small’s release has several implications:
- Efficiency: Demonstrates that smaller, efficient models can match larger ones
- Accessibility: Open weights and efficient architecture make it accessible to developers
- Cost: Lower inference costs make it attractive for high-volume applications
- Innovation: Shows that model architecture matters as much as model size
The release challenges the assumption that bigger is always better in AI. By optimizing architecture and training, smaller models can achieve competitive performance at a fraction of the cost.
Context
Inkling-Small is the second model from Thinking Machines Lab, founded by former OpenAI CTO Mira Murati. The company’s first model, Inkling (975B parameters), was released under Apache 2.0 in July 2026.
The release reflects a broader trend toward efficient, open-weight models that prioritize practical utility over benchmark scores. As the AI market matures, the ability to deliver competitive performance at lower cost becomes increasingly important.