Inkling - A multimodal foundational model from Thinking Machines Lab

What is Inkling?

Inkling is an open-weighted multimodal foundational model from Thinking Machines Lab. The model natively supports text, image, and audio input, employs a hybrid expert architecture, and achieves cross-modal inference with 41 billion activation parameters. Developers can download weights and fine-tune them as needed through the Tinker platform. Unlike models that aim for top scores in a single benchmark, Inkling strives for balanced performance across a wide range of tasks, including agent-based, programming, vision, and audio tasks, making it a versatile backend inference engine that can be embedded in different workflows.

Inkling's main functions

• Multimodal native understandingIt can simultaneously process text, image, and audio input, and supports speech-to-text, visual question answering, chart reasoning, and long audio analysis.
• Intelligent agents and tool callsIt has built-in multi-framework tool capabilities, enabling browser operations, code generation, terminal tasks, and multi-round collaborative development.
• Controllable reasoning strengthIt supports adjusting the inference depth (effort parameter), allowing for flexible trade-offs between token cost and performance, while maintaining high efficiency in low-latency scenarios.
• Long context processingIt supports up to 1 million token contexts, making it suitable for long document analysis, multi-turn dialogues, and complex workflows.
• Open weights and fine-tuningThe full weights are made available through Hugging Face, and together with the Tinker platform, they provide end-to-end fine-tuning and self-training capabilities.

Inkling's technical principles

• MoE Sparse Architecture: 66-layer decoder-only Transformer, with 6 experts activated per token among 256 experts + 2 shared experts, total parameters 975 billion, and activation parameters 41 billion.
• Encoderless multimodalThe image is encoded using four layers of hMLP in 40×40 pixel blocks, and the audio is input as discrete tokens of the dMel spectrogram. All modalities are projected to a shared latent space for unified decoding.
• Hybrid attention mechanismAlternating between local and global attention layers balances long contextual memory and computational efficiency.
• Quantitative deploymentSupports BF16, MXFP8 and NVFP4 precision. After NVFP4 quantization, it only requires 600GB of video memory to run.
• Training dataPre-trained on text, image, audio, and video data containing 45 trillion tokens, the data was deduplicated, filtered, and synthesized for enhancement.

Follow us on WeChat and reply with "open source",join inAI open source project discussion group

How to use Inkling

• Online experienceAccess the Inkling Playground (Tinker console) https://tinker.thinkingmachines.ai/playground to interact directly with the model and test its style and capabilities.
• API callsAccess the API through the Tinker platform or a third-party inference service provider, and use it. tml-renderers Package integration is quick.
• Local deploymentDownload weights from Hugging Face and deploy using the SGLang, vLLM, TokenSpeed, Unsloth, or Hugging Face framework; BF16 requires 2TB of VRAM, and NVFP4 Quantized Edition requires 600GB.
• Fine-tuning customizationUpload data to the Tinker platform, write fine-tuning tasks, and run training to allow the model to participate in its own iterative optimization.
• Reasoning intensity adjustmentThe effort parameter controls the number of tokens generated, matching cost and performance requirements.

Inkling's core advantages

• High-efficiency inference curveIn tests such as Terminal Bench, Nemotron 3 Ultra achieved equivalent performance with one-third of the tokens, significantly reducing the cost of long processes.
• Balanced general abilitiesIt does not overfit to a single domain, but maintains competitiveness in dimensions such as intelligent agents, programming, vision, audio, and instruction compliance.
• Native MultimodalThe audio and visual components are trained from scratch, without post-processing. Speech and image understanding are among the top tier of open-weight models.
• Safety and controllabilityBuilt-in FORTRESS security protection with a low false rejection rate; supports the overlay of downstream review tools such as Llama Guard to form a defense-in-depth system.

Inkling's project address

• Project official websitehttps://thinkingmachines.ai/news/introducing-inkling/
• HuggingFace model libraryhttps://huggingface.co/thinkingmachines/Inkling
• Technical Papershttps://thinkingmachines.ai/model-card/inkling/

Comparison of Inkling's similar products

Applications of Inkling

• Enterprise Intelligent Agent CentralAs a backend inference model, it drives the collaboration of multiple tools such as customer service, data analysis, and code generation into an intelligent agent system, which is fine-tuned to adapt to the internal workflow of the enterprise.
• Multimodal content moderationIt simultaneously analyzes text, images, and audio content for efficient compliance review and risk identification on social media and live streaming platforms.
• Interactive Educational AssistantBased on voice and visual input, it provides students with real-time problem-solving, diagrammatic explanations, and programming tutoring, and supports teaching material contexts with millions of tokens.
• Low-latency programming assistantIn the low effort parameter mode, code completion, terminal command generation, and lightweight bug fixing are completed with extremely low latency, and an IDE plugin is embedded.
• Long document legal/financial analysisUsing a million token contexts, perform cross-page correlation reasoning on contracts, research reports, and financial statements to extract key terms and risk points.