Thinking Machines redefines open-source AI with Inkling-Small launch
Quarter the size, nearly the same performance: Inkling-Small sets new industry standards.
1Next-gen open source AI model
Thinking Machines unveiled Inkling-Small, a new open-source AI model with 276 billion parameters. Licensed under Apache 2.0, it maintains just 2.4% performance gap compared to its predecessor Inkling despite being a quarter the size. The multimodal model supports inputs up to 1 million tokens. Complete model weights are available on Hugging Face, with fine-tuning support through Tinker API.
2Performance benchmarks
Inkling-Small scored 40 on Artificial Analysis Intelligence Index versus Inkling's 41. It achieved 80.2% on SWE-bench Verified and 64.7% on Terminal Bench 2.1, outperforming its larger sibling in some tests. However, it lagged in specialized tasks like financial transactions (15.5% vs 23.7%). Experts highlight its balanced performance across different applications while noting domain-specific limitations.
3Architecture insights
The model employs sparse Mixture-of-Experts architecture. Its 42-layer decoder selects 6 out of 256 experts per token, maintaining 12B active parameters despite 276B total. This design preserves capability while reducing compute costs. Text, image and audio inputs are processed through shared representations. Thinking Machines provides fine-tuning capabilities via Tinker API, enabling customization for specific use cases.
4Enterprise advantages
Smaller size significantly reduces deployment costs. Launch pricing offers 50% discount at $0.58 per million input tokens and $1.44 per output tokens. Requiring 3.5x less resources than Inkling, it lowers GPU demands. However, minimum 600GB GPU memory requirement makes it unsuitable for consumer hardware. The model targets corporate servers and cloud solutions, particularly appealing for data-sensitive organizations.
5Hardware requirements
Inkling-Small needs 4x NVIDIA B300 or 8x H200 GPUs. NVFP4 quantization reduces VRAM needs to 180GB, enabling W4A4 operation on single B300. These specs exclude laptops and desktops, positioning it for enterprise infrastructure. Thinking Machines emphasizes its suitability for companies prioritizing data control and customization over maximum performance.
6Application areas
The model suits coding assistants, chatbots, and document analysis. It supports multimodal workflows and retrieval-augmented generation. Variable reasoning effort allows compute adjustment based on task difficulty, giving engineering teams cost-performance flexibility. Thinking Machines recommends human oversight for high-stakes factual tasks, acknowledging its limitations in specialized domains.
7Industry impact
Experts recognize Inkling-Small as a milestone for open-source AI. Its size-performance balance improves accessibility for mid-sized companies. Cost advantages could democratize AI product development, though some caution about limitations in niche applications. The launch may accelerate similar compact model development across the industry.
8Future outlook
Thinking Machines' move highlights industry focus on size-efficiency optimization. More companies will likely develop similar models, with open-source approach fostering community contributions. Analysts predict adoption in data-sensitive sectors. Reduced AI infrastructure costs could accelerate innovation cycles, though maintaining quality in specialized domains remains challenging.