Thinking Machines Releases Inkling-Small Model, Capable of Single Node Operation and Emotion Detection
Thinking Machines has released Inkling-Small: with a total of 276B parameters and 12B active parameters, it operates faster and matches or surpasses its 975B counterpart in multiple benchmarks.
The team has integrated it into a voice-to-voice system, where audio is directly input into the model and responded to using faster-Qwen3TTS, achieving a latency of less than 500 milliseconds on 8 RTX Pro 6000 Blackwell GPUs.
The standard Inkling requires 2TB of VRAM, while this version can operate on a single node and is capable of detecting tone and emotion.
Source: Public Information
ABAB AI Insight
Thinking Machines was founded by former OpenAI CTO Mira Murati and has previously open-sourced the Inkling large model. This release offers a lighter variant to balance performance and deployment costs.
On the capital front, open-sourcing weights lowers the inference barrier, attracting developers to experiment with multimodal agents on single nodes or consumer-grade clusters, accelerating the transition from lab to practical application.
Similar to DeepSeek and other MoE models with small active parameters, the current open-source frontier is in the "large total parameters + small active" efficiency optimization phase.
This essentially represents a technological substitution, where smaller active parameters significantly reduce hardware barriers while maintaining capability, promoting the proliferation of physical and voice agents.
ABAB News · Cognitive Laws
Active parameters determine real deployment costs
Small models matching large models is an efficiency victory
Tone perception allows interaction to move from text to emotion