New Model: Inkling by Thinking Machine on Hugging Face

2026-07-16 · Source: HuggingFace · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Advanced, extended

Summary

Thinking Machines has launched "Inkling," a new multimodal model on Hugging Face Hub, featuring nearly one trillion parameters and Apache 2.0 licensing. This model uniquely handles image, text, and audio inputs, demonstrating strong multimodal reasoning capabilities, comparable to Gemini Pro 3.1. It offers a 1 million context window and is available in BF16 and NVFP4 formats, alongside a speculative multi-token prediction drafter for enhanced speed. Architecturally, Inkling employs a distinctive relative attention mechanism and SC convolutions, with shallow multimodal modules. The model, internally sized by depth (D66), supports zero-shot inference via Transformers, Llama CPP, SG lang, and VLM. Demos showcase its ability in visual and audio reasoning, code generation, and data extraction, with specific "reasoning efforts" settings recommended for optimal performance. It also emphasizes post-training and personalization, integrating with platforms like Tinker and OpenMend.

Key takeaway

For MLOps Engineers evaluating new multimodal foundation models, Inkling presents a compelling open-source option. Its nearly one trillion parameters, 1 million context window, and native image, text, and audio processing capabilities make it suitable for complex reasoning tasks. You should explore its NVFP4 variants and Llama CPP integrations for efficient deployment, and consider its post-training focus with platforms like Tinker for tailored applications.

Key insights

Multimodal models can achieve robust reasoning across image, text, and audio with simplified encoder architectures.

Principles

• Open-weight models offer significant customization potential beyond initial API outputs.
• Relative attention can effectively replace absolute positional embeddings in Transformer architectures.
• Model architecture can prioritize depth and sparsity over raw parameter count for scaling.

In practice

• Test Inkling via Hugging Face inference providers for multimodal reasoning tasks.
• Explore post-training with Tinker or OpenMend's echo algorithm for model adaptation.

Topics

• Multimodal Foundation Models
• Hugging Face Hub
• Model Architecture
• Model Quantization
• Post-training
• Inference Optimization

Best for: AI Architect, AI Engineer, NLP Engineer, AI Scientist, Machine Learning Engineer, MLOps Engineer

Related on AIssential

• not much happened today — Inkling's launch emphasizes an open-weight, customizable multimodal foundation model designed for practical application and ecosystem integration over pure benchmark leadership.
• Multimodal Max — Multimodal Max dynamically routes to specialized models, achieving superior performance and latency across diverse tasks.
• Seeing Before Reasoning: Decoupling Perception and Reasoning for Shortcut-Resilient Multimodal On-Policy Self-Distillation — Decoupling perception and reasoning in MLLMs prevents shortcut learning by ensuring visual grounding.
• Together AI brings Thinking Machines Lab’s new model Inkling on day 0 — Inkling unifies multimodal inputs and offers controllable reasoning via a novel MoE architecture for broad task versatility.
• Thinking Machines Lab offers enterprises a US alternative in open-weight AI — Inkling offers Western enterprises a US-developed, open-weight, multimodal AI model for domain-specific adaptation.
• Thinking Machines Lab Drops Its First Model — Thinking Machines Lab launched Inkling, a large open-source multimodal model, challenging established AI players with a vision for decentralized AI.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by HuggingFace.