Inkling Open-Weight Multimodal Model Debuts with 975B Parameters

TL;DR: Thinking Machines released Inkling, a 975 billion-parameter open-weight multimodal model that activates about 41 billion parameters per task and supports up to a 1-million token context. It matches Nvidia’s Nemotron 3 Ultra coding scores while using roughly a third of the tokens, and is available for download and fine-tuning via the Tinker platform and HuggingFace, positioning it as a Western alternative to Chinese open-source models.

Key facts

• Inkling was unveiled on July 15, 2026 by Thinking Machines Lab, founded by former OpenAI CTO Mira Murati. [2][3][4][5]
• Inkling is an open-weight Mixture-of-Experts transformer with 975 billion total parameters and about 41 billion active parameters per task. [2][3][4][5][6]
• The model was pretrained on roughly 45 trillion multimodal tokens covering text, images, audio and video. [3][4][5][6]
• Inkling supports a context window of up to 1 million tokens, though the hosted platform caps usage at 256 K tokens. [3][4]
• Benchmark scores include 77.6% on SWE-bench Verified and 97.1% on AIME 2026, with top-tier results on audio suites VoiceBench, MMAU and AudioMC. [2][4][6]
• Users can adjust a “thinking effort” knob to trade speed for accuracy; the company claims coding performance matches Nvidia’s Nemotron 3 Ultra while consuming about one-third as many tokens. [4][5][6]
• Inkling can be downloaded and fine-tuned via the Tinker platform, is listed on HuggingFace, and early customers such as Bridgewater Associates are already using it. Revenue comes from Tinker services, not the model itself. [2][3][4][5][6]

Introduction

Thinking Machines Lab announced Inkling on July 15, 2026, presenting it as the first open-weight, multimodal Mixture-of-Experts model from a Western AI startup. Founded by former OpenAI CTO Mira Murati, the company aims to give enterprises a flexible alternative to the dominant Chinese open-source offerings.

Technical specifications

Inkling is built as a Mixture-of-Experts transformer, meaning that a large pool of expert sub-networks is sparsely activated for each input. The public specification lists 975 billion total parameters, but only about 41 billion parameters are active during inference. This sparsity enables the model to process a context window of up to 1 million tokens, although the hosted service caps usage at 256 K tokens to manage compute load.

The model was pretrained on an estimated 45 trillion multimodal tokens that span text, images, audio, and video, giving it a broad knowledge base current through April 2026. Inkling accepts text, image and audio inputs, while its outputs are presently limited to text, covering natural language, code, and structured data.

Benchmark performance

In internal testing, Inkling achieved 77.6% on SWE-bench Verified, a coding benchmark, and 97.1% on AIME 2026, an advanced AI-driven evaluation suite. Audio-focused benchmarks such as VoiceBench, MMAU and AudioMC also placed Inkling among the top performers, indicating strong cross-modal capabilities.

A notable claim is that Inkling matches the coding benchmark scores of Nvidia’s Nemotron 3 Ultra while consuming roughly one-third as many tokens. This efficiency is attributed to the model’s ability to deactivate most of its parameters for a given task.

Customization and pricing model

Inkling introduces a “thinking effort” knob that lets users trade speed for accuracy. Raising the knob increases the number of experts consulted, improving result quality at the cost of longer latency. This feature is designed for enterprises that need to balance real-time responsiveness with precision.

Instead of charging per-token API fees, Thinking Machines monetizes the Tinker platform, a suite that hosts the model, provides fine-tuning tools, and offers the interactive Inkling Playground console. Early adopters such as Bridgewater Associates have already signed up for the service, indicating market interest from the financial sector.

Strategic positioning against Chinese models

The launch is framed as a Western counter-weight to Chinese open-source models like Alibaba’s Qwen. By releasing an open-weight model, Thinking Machines hopes to shift AI spend toward in-house infrastructure where companies can tailor the model to proprietary data and compliance requirements. The strategy also aligns with broader concerns about supply-chain dependence on non-Western AI providers.

Availability and ecosystem integration

Inkling’s weights are downloadable and can be fine-tuned via the Tinker platform. The model is also listed on HuggingFace, making it accessible to the broader developer community. Partner integrations extend its reach into cloud marketplaces and specialized AI services, allowing developers to experiment without setting up their own hardware.

Community reaction

The AI community has highlighted Inkling’s thinking effort knob as a novel way to give developers direct control over the speed-accuracy trade-off, a feature rarely exposed in large-scale models. Observers note that the combination of open-weight access, multimodal capability, and a large context window could spur new research in long-form reasoning and multimodal generation. [1]

Outlook

Inkling’s release marks a significant step for Western AI startups seeking to compete in the open-weight space. Its blend of massive scale, sparsity-driven efficiency, and developer-friendly tooling may encourage more enterprises to move away from per-token API pricing toward self-hosted, customizable foundations. How quickly the ecosystem adopts Inkling—particularly in comparison with entrenched Chinese models—will shape the next wave of AI infrastructure decisions.

All data and benchmark figures are taken from Thinking Machines’ public announcements and partner reports.

Sources

1. What Thinking Machines’ Inkling Is Really Like, Part I (semaphore.substack.com) — 2026-07-20