Thinking Machines Unveils Inkling: A New Multimodal MoE Model for Developers

What Changed

Thinking Machines has officially released Inkling, a new general-purpose multimodal model. Inkling distinguishes itself by accepting text, image, and audio inputs, and producing text outputs. The model is released with open weights, targeting developers who are building AI-powered applications such as agentic systems, coding assistants, chatbots, and retrieval-augmented generation (RAG) systems. It is also suitable for general conversational use and instruction-following tasks.

Inkling supports local deployment through integration with several open-source libraries, including SGLang, vLLM, TokenSpeed, Unsloth, and Hugging Face's Transformers library. API access is also provided via third-party inference providers. The model is intended for use primarily in English but possesses general multilingual capabilities.

Technical Details

Inkling is architected as a 66-layer decoder-only transformer with a sparse Mixture-of-Experts (MoE) feed-forward backbone. This architecture routes each token to 6 out of 256 experts, in addition to 2 shared experts that are active on every token. The attention mechanism is a hybrid of local and global layers.

The model is natively multimodal, processing images and video through a hierarchical patch encoder and audio via discrete token encoding. All modalities are projected into a shared hidden space and processed jointly by the decoder. Inkling has 975 billion total parameters, with 41 billion active parameters. It supports BF16 and NVFP4 numerics.

Input modalities include UTF-8 encoded text, pixel-based image formats (ideally between 40px and 4096px per dimension), and WAV audio sampled at 16kHz (ideally under 20 minutes in length). The model generates output as UTF-8 encoded text.

Training data for Inkling was sourced from publicly available content, third-party acquisitions, and synthetically generated or augmented datasets. This data includes a broad variety of text, images, audio, and video, undergoing processes such as deduplication and filtering for quality and safety objectives.

Benchmark Analysis

Evaluations for Inkling were conducted with results reported at effort=0.99, with comparison scores generated on July 14, 2026. Inkling was benchmarked against both open-weights models (Nemotron 3 Ultra, Kimi K2.5, Kimi K2.6, GLM 5.2, DeepSeek V4 Pro) and closed-weights models (Gemini 3.1 Pro, Claude Fable 5, GPT 5.6 Sol) across various categories.

In Reasoning tasks, Inkling achieved 29.7% on HLE (text only) and 46.0% on HLE (with tools). It scored 97.1% on AIME 2026 and 87.2% on GPQA Diamond.

For Agentic (coding) tasks, Inkling demonstrated 77.6% on SWEBench Verified, 54.3% on SWEBench Pro (Public), and 63.8 on Terminal Bench 2.1 (Best Harness). Its GDPVal-AA v2 score was 1233.

In Agentic (general) evaluations, Inkling scored 74.1% on MCP Atlas and 23.7% on Tau 3 Banking.

Factuality benchmarks showed Inkling at 77.1% on BrowseComp (w/ Ctx), 43.9% on SimpleQA Verified, and 1.0% on AA Omniscience.

For Chat capabilities, Inkling achieved 79.8% on IFBench and 88.7% on Global-MMLU-Lite.

Vision tasks saw Inkling score 73.3% on MMMU Pro (Standard 10), 78.1% on Charxiv RQ, and 82.0% on Charxiv RQ (with python).

In Audio evaluations, Inkling achieved 56.6% on Audio MC, 77.2% on MMAU, and 91.4% on VoiceBench.

Safety evaluations indicated Inkling scored 78.0% on FORTRESS (Adversarial), 95.9% on FORTRESS (Benign), and 98.6% on StrongREJECT.

Developer Implications

Inkling's open-weights release and multimodal capabilities provide developers with a flexible foundation for a wide array of AI applications. The model's support for various local deployment libraries (SGLang, vLLM, TokenSpeed, Unsloth, Hugging Face) simplifies integration into existing development workflows and allows for fine-tuning and research. Its design for agentic and tool-use systems, coding assistants, and chatbots suggests direct applicability in enhancing automation and interactive AI experiences.

The model's multilingual capabilities, while primarily English-focused, indicate potential for global application development. Developers should consider the provided documentation and Tinker Cookbook for getting started. The emphasis on safety evaluations and the recommendation for defense-in-depth strategies, including layering moderation tools like Llama Guard, highlight the need for responsible deployment, especially in consumer-facing or high-traffic scenarios.

Bottom Line

Inkling represents a new multimodal MoE model from Thinking Machines, offering open weights and broad input/output modality support. Its architecture, featuring a sparse Mixture-of-Experts, aims to provide a general-purpose foundation for developers building advanced AI applications. While benchmarks show competitive performance across various tasks, developers are advised to conduct their own evaluations for specific use cases and implement additional safeguards to address potential limitations such as hallucination, instruction-following failures, and biases inherent in large-scale models.