Open-weights multimodal model with 975 billion parameters trained on 45 trillion tokens across text, image, audio and video.

Key facts

Overview

Inkling is an open-weights multimodal large language model released by Thinking Machines in July 2026, consisting of 975 billion total parameters pretrained on 45 trillion tokens spanning text, images, audio and video modalities. Built in-house over 1.5 years largely out of public view, it represents a direct open-source alternative to proprietary offerings from major US AI labs and Chinese competitors. The weights are publicly available, permitting local deployment without API calls or service subscriptions. Unlike closed models such as GPT-5.5 or Gemini 2.5 Pro, which charge per token and centrally control model updates, Inkling trades immediate vendor support and extensive public benchmarking for transparency, reproducibility and full user control over inference behaviour. Early performance data is limited; the model has not yet been extensively evaluated against contemporary closed-source baselines on standardised benchmarks, making direct capability comparison speculative. Typical use cases include custom domain fine-tuning, private inference in regulated sectors, and research into open multimodal architectures. Organisations with substantial GPU or TPU capacity and in-house ML expertise can deploy Inkling entirely on private infrastructure, eliminating recurring API costs. Conversely, teams lacking dedicated ML operations capacity or needing service-level guarantees, technical support and rapid model updates will find closed-source subscription models more practical. The broader open-source ecosystem—particularly Hugging Face and community fine-tunes—may offer more specialized variants tailored to specific tasks, and inference frameworks like Ollama are beginning to package open models for ease of deployment. Inkling's competitive positioning depends largely on how rapidly the community develops fine-tuned variants and integration libraries.

Who it is for

• Best for: Organisations and researchers seeking open, locally-deployable multimodal AI without API dependencies..
• Avoid if: You need production-grade guarantees, fine-tuned performance on narrow tasks, or prefer vendor support and service-level agreements..

Who this is for

Machine learning engineers, research scientists, and open-source advocates building custom AI systems. Also suitable for enterprises with dedicated ML operations teams seeking to avoid vendor lock-in and maintain full control over model weights and inference infrastructure.

Who should skip this

Teams requiring immediate, turnkey solutions with vendor backing; startups without dedicated ML infrastructure; organisations subject to strict audit and compliance regimes that demand traceable, supported model provenance.

Verdict

Inkling is a significant open-weights release that expands the available landscape of publicly auditable, locally-deployable multimodal AI, appealing to researchers and enterprises prioritising control and cost efficiency over turnkey support. However, early-stage maturity, limited public benchmarking, and substantial infrastructure requirements restrict its utility to organisations with dedicated ML teams. For most users, established closed-source services remain more practical in the near term, though Inkling's open nature positions it as a long-term hedge against vendor dependency.

Inkling FAQ

What is Inkling? +

Inkling is an open-weights multimodal large language model released by Thinking Machines in July 2026, consisting of 975 billion total parameters pretrained on 45 trillion tokens spanning text, images, audio and video modalities. Built in-house over 1.5 years largely out of public view, it represents a direct open-source alternative to proprietary offerings from major US AI labs and Chinese competitors. The weights are publicly available, permitting local deployment without API calls or service subscriptions. Unlike closed models such as GPT-5.5 or Gemini 2.5 Pro, which charge per token and centrally control model updates, Inkling trades immediate vendor support and extensive public benchmarking for transparency, reproducibility and full user control over inference behaviour. Early performance data is limited; the model has not yet been extensively evaluated against contemporary closed-source baselines on standardised benchmarks, making direct capability comparison speculative. Typical use cases include custom domain fine-tuning, private inference in regulated sectors, and research into open multimodal architectures. Organisations with substantial GPU or TPU capacity and in-house ML expertise can deploy Inkling entirely on private infrastructure, eliminating recurring API costs. Conversely, teams lacking dedicated ML operations capacity or needing service-level guarantees, technical support and rapid model updates will find closed-source subscription models more practical. The broader open-source ecosystem—particularly Hugging Face and community fine-tunes—may offer more specialized variants tailored to specific tasks, and inference frameworks like Ollama are beginning to package open models for ease of deployment. Inkling's competitive positioning depends largely on how rapidly the community develops fine-tuned variants and integration libraries.

How much does Inkling cost? +

Inkling pricing: Open-weights (free to download and run locally). Always confirm current pricing on the official site, as plans change.

Does Inkling have a free tier? +

Yes. Inkling offers a free plan or free credits you can use to evaluate it.

What is Inkling best for? +

Organisations and researchers seeking open, locally-deployable multimodal AI without API dependencies..

When should you avoid Inkling? +

Avoid Inkling if: You need production-grade guarantees, fine-tuned performance on narrow tasks, or prefer vendor support and service-level agreements..

What are the main pros of Inkling? +

Openly available weights enable local deployment without reliance on proprietary APIs; Multimodal training across text, image, audio and video supports diverse reasoning tasks; Developed after 1.5 years of private research, indicating substantial engineering depth.

What are the main cons of Inkling? +

Limited public benchmarking data compared to established closed-source models like GPT-5.5 or Gemini 2.5 Pro; Requires significant local compute resources; accessibility for resource-constrained users remains unclear; Early-stage adoption means limited community documentation and integration patterns.

Does Inkling have an affiliate program? +

No public affiliate program is listed for Inkling at the time of review.

How is Inkling rated? +

WireTensors rates Inkling 3.8 out of 5, based on capability, value, and fit for its intended use case.

What category does Inkling fall under? +

Inkling is categorised under coding on WireTensors.

When was this Inkling review last verified? +

This review was last verified on 2026-07-16 against the vendor's official site.

Sources

• Inkling — official website — verified 2026-07-16