- Published on
Thinking Machines debuts Inkling Small open source AI model nearing performance of predecessor at about 1/4 size
- Authors
- Name
- DeepHorus
Thinking Machines has unveiled Inkling Small, a new open source AI model that delivers nearly the performance of its flagship Inkling at roughly one-quarter the size, just two weeks after the company released its first model. The startup, led by former OpenAI chief technology officer Mira Murati, introduced the 276-billion-parameter multimodal reasoning model with a permissive Apache 2.0 license. It accepts text, image, and audio inputs, produces text, and supports a context window of up to one million tokens.
On the third-party Artificial Analysis Intelligence Index, Inkling Small scored 40, within one point of Inkling's 41, despite the flagship having 975 billion total parameters. This is notable because no open-weight model at Inkling Small's size or smaller has scored higher on the index. The smaller model also exceeds its predecessor on several benchmarks, including 80.2% on SWE-bench Verified versus 77.6%, and 64.7% on Terminal Bench 2.1 versus 63.8%.
For enterprises, the appeal is not simply size. Inkling Small uses only 12 billion active parameters per token, compared with 41 billion for Inkling, while preserving much of the flagship's coding, reasoning, and multimodal performance. That translates into lower compute requirements, inference costs, and deployment footprint. It remains far too large for a laptop, but it is materially easier to operate than the 3.5x larger flagship, making it a good fit for companies with modest GPU resources.
The performance gains are not universal. Inkling Small trails on factual knowledge and some agentic tasks, scoring 15.5% on τ³-Banking versus 23.7% for Inkling, with a negative AA Omniscience score. That tradeoff matters for organizations using it for high-stakes factual tasks, which will still need retrieval, verification, and human review. The model is better suited for coding assistants, tool-use systems, retrieval-augmented generation, document analysis, and multimodal workflows.
The architecture behind Inkling Small is a sparse Mixture-of-Experts model, according to its model card. The 42-layer decoder routes each token to six of 256 specialized experts, plus two shared experts that remain active for every token. This explains the gap between the 276 billion total parameters and the 12 billion active parameters, allowing a large pool of learned capacity with only a fraction activated at a time.
Thinking Machines has released the full weights on Hugging Face and added fine-tuning support through its Tinker API. At launch, it is offering a limited-time 50% discount, bringing prefill pricing to $0.58 per million tokens, sampled output to $1.44 per million, and training to $1.73 per million, with cached prefill at $0.116 per million. A 256K-context variant is available at higher rates.
