Inkling - A multimodal foundation model from Thinking Machines Lab

What is Inkling

Inkling is an open-weight multimodal foundation model from Thinking Machines Lab. The model natively supports text, image, and audio input, adopts a mixture-of-experts architecture, and achieves cross-modal reasoning with 41 billion activated parameters, allowing developers to download weights and fine-tune on demand via the Tinker platform. Unlike models that chase single-benchmark championships, Inkling pursues balanced performance across a wide range of tasks such as agents, coding, vision, and audio, becoming a backend inference engine that can be embedded into different workflows.

Inkling's main features

• Native multimodal understanding: Simultaneously processes text, image, and audio input, supporting speech transcription, visual question answering, chart reasoning, and long audio analysis.
• Agents and tool calling: Built-in multi-framework tool use capabilities, enabling browser operations, code generation, terminal tasks, and multi-turn collaborative development.
• Controllable reasoning intensity: Supports adjusting reasoning depth (effort parameter), flexibly balancing token cost and performance while remaining efficient in low-latency scenarios.
• Ultra-long context processing: Supports up to 1 million tokens of context, suitable for long document analysis, multi-turn conversations, and complex workflows.
• Open weights and fine-tuning: Complete weights are open-sourced via Hugging Face, with the Tinker platform providing end-to-end fine-tuning and self-training capabilities.

Inkling's technical principles

• MoE sparse architecture: A 66-layer decoder-only Transformer, with 6 experts activated per token out of 256 experts + 2 shared experts, 975 billion total parameters and 41 billion activated parameters.
• Encoder-free multimodality: Images are encoded via 40×40 pixel patches through a four-layer hMLP, audio is input as discrete tokens from dMel spectrograms, and all modalities are projected into a shared latent space for unified decoding.
• Hybrid attention mechanism: Local and global attention layers alternate, balancing long-context memory and computational efficiency.
• Quantized deployment: Supports BF16, MXFP8, and NVFP4 precision; after NVFP4 quantization, only 600GB of VRAM is needed to run.
• Training data: Pre-trained on 45 trillion tokens of text, image, audio, and video data, with data deduplicated, filtered, and synthetically augmented.

Follow WeChat and reply “open source” to join theAI open-source project communication group

How to use Inkling

• Online experience: Visit the Inkling Playground (Tinker console)https://tinker.thinkingmachines.ai/playground直接与模型交互，测试风格与能力。
• API calls: Access the API through the Tinker platform or third-party inference providers, using the tml-renderers package for quick integration.
• Local deployment: Download weights from Hugging Face and deploy using SGLang, vLLM, TokenSpeed, Unsloth, or the Hugging Face framework; BF16 requires 2TB of VRAM, while the NVFP4 quantized version requires 600GB.
• Fine-tuning and customization: Upload data on the Tinker platform, write fine-tuning tasks, and run training, allowing the model to participate in its own iterative optimization.
• Reasoning intensity adjustment: Control the number of generated tokens via the effort parameter to match cost and performance needs.

Inkling's core advantages

• Efficient inference curve: In tests such as Terminal Bench, it achieves comparable performance using one-third of the tokens of Nemotron 3 Ultra, significantly reducing the cost of long workflows.
• Balanced generalist capabilities: It does not overfit to a single domain, remaining competitive simultaneously across dimensions such as agents, coding, vision, audio, and instruction following.
• Native multimodal: The audio and vision components are trained from scratch, not bolted on afterward, and speech and image understanding are in the first tier among open-weight models.
• Safety and controllability: Built-in FORTRESS safety protections with a low false-refusal rate; supports stacking downstream moderation tools such as Llama Guard to form defense in depth.

Inkling's project address

• Project website：https://thinkingmachines.ai/news/introducing-inkling/
• HuggingFace model repository：https://huggingface.co/thinkingmachines/Inkling
• Technical paper：https://thinkingmachines.ai/model-card/inkling/

Comparison of Inkling with similar competing products

Inkling's application scenarios

• Enterprise agent hub: As a backend inference model, it drives agent systems that coordinate multiple tools such as customer service, data analysis, and code generation, adapting to enterprise internal workflows through fine-tuning.
• Multimodal content moderation: It analyzes text, image, and audio content simultaneously, for efficient compliance review and risk identification on social platforms and livestreaming platforms.
• Interactive education assistant: Based on speech and visual input, it provides students with real-time problem solving, chart explanation, and programming tutoring, supporting a million-token long textbook context.
• Low-latency coding assistant: In low effort parameter mode, it completes code completion, terminal command generation, and lightweight bug fixes with extremely low latency, embedded in IDE plugins.
• Long-document legal/financial analysis: Using a million-token context, it performs cross-page associative reasoning on contracts, research reports, and financial statements, extracting key clauses and risk points.