Thinking Machines Lab releases Inkling Small multimodal AI model

Inkling Small Thinking Machines Lab open‑weight AI model benchmark performance Apache 2.0

Thinking Machines Lab, the AI laboratory founded by former OpenAI CTO Mira Murati, made the full weights of its new multimodal model Inkling Small publicly available on July 30, 2026. The model contains 276 billion total parameters, of which 12 billion are active, making it roughly one‑quarter the size of the earlier Inkling model (975 billion total, 41 billion active).

Inkling Small scores 40 on the Artificial Analysis Intelligence Index, one point below Inkling’s 41, and, according to the same analysis, no other open‑weight model of equal or smaller size scores higher. It surpasses its larger sibling on several coding and reasoning benchmarks, such as Humanity’s Last Exam (32 % vs. 30 %) and GPQA Diamond (89 % vs. 87 %). The model is also markedly more token‑efficient, averaging 24 K output tokens per task compared with 45 K for DeepSeek V4 Flash and 78 K for GPT‑5.

The system processes text, image, and speech inputs, offers a context window of 256 K tokens (the Japanese release notes a 1 million‑token context length), and is distributed under an Apache 2.0 license on Hugging Face. Users can fine‑tune the model via the Tinker Playground and run inference with engines such as SGLang, vLLM, TokenSpeed, and Unsloth.

Entities

Inkling Small · Mira Murati · Thinking Machines Lab

Claims

What the coverage asserts, and how well corroborated each claim is across sources.

[● 3 SOURCES] Inkling Small's weights are released under an Apache 2.0 license on Hugging Face. tech-noisy.com · the-decoder.de · the-decoder.com
[● 2 SOURCES] Inkling Small scores 40 on the Artificial Analysis Intelligence Index, one point below Inkling's 41. the-decoder.de · the-decoder.com
[● 2 SOURCES] Inkling Small averages 24 K output tokens per task, compared with 45 K for DeepSeek V4 Flash and 78 K for GPT‑5. the-decoder.de · the-decoder.com
[○ 1 SOURCE] Inkling Small is about one‑quarter the size of Inkling (Inkling 975 B total, 41 B active). tech-noisy.com
[● 3 SOURCES] Inkling Small has 276 billion total parameters and 12 billion active parameters. tech-noisy.com · the-decoder.de · the-decoder.com
[○ 1 SOURCE] Thinking Machines Lab released Inkling Small on July 30, 2026. tech-noisy.com
[● 2 SOURCES] Inkling Small supports text, image, and speech inputs and has a 256 K-token context window. the-decoder.de · the-decoder.com
[● 2 SOURCES] Inkling Small outperforms Inkling on Humanity's Last Exam (32 % vs. 30 %) and GPQA Diamond (89 % vs. 87 %). the-decoder.de · the-decoder.com

Sources

Thinking Machines setzt mit zweitem Modell Inkling Small auf Effizienz statt Größe [the-decoder.de]

12 days ago

Inkling-Small - 総276B・アクティブ12Bの構成とベンチマーク・入手方法 [tech-noisy.com]

12 days ago

Thinking Machines bets on efficiency over size with its second model, Inkling Small [the-decoder.com]

11 days ago