Thinking Machines Releases Inkling: A 975B Open-Weight Multimodal AI

Thinking Machines Lab Releases Inkling on 15 July 2026

On 15 July 2026, Thinking Machines Lab — the AI company founded by Mira Murati, former Chief Technology Officer of OpenAI — released Inkling, its first in-house AI model. Inkling is a 975-billion-parameter sparse Mixture-of-Experts model with 41 billion active parameters, trained on 45 trillion tokens spanning text, images, audio, and video. The weights are available on Hugging Face under an Apache 2.0 licence, alongside an NVFP4 quantised checkpoint tuned for NVIDIA Blackwell hardware. Thinking Machines is also previewing Inkling-Small, a 276-billion-parameter variant with 12 billion active parameters. The company explicitly positioned Inkling not as the world's strongest model but as the best open-weight starting point for organisations that want to customise AI for their specific use cases.

Architecture: 975B Parameters, 41B Active

Inkling's sparse Mixture-of-Experts architecture activates only 41 billion of its 975 billion total parameters for any given forward pass, keeping inference costs manageable despite the model's large total size. The model supports a one-million-token context window and reasons natively across text, images, and audio — no separate modality pipeline is required. The NVFP4 checkpoint on Hugging Face is optimised for NVIDIA's Blackwell-generation GPUs, reducing the memory footprint for self-hosted inference on current hardware. Inkling-Small, at 276 billion total and 12 billion active parameters, matched or surpassed the flagship on several benchmarks in Thinking Machines' internal testing, providing a lower-infrastructure alternative for teams that cannot provision the hardware required for the larger model.

Benchmarks: Leading Open-Weight Results from a US Lab

Inkling's benchmark results place it at the top of publicly available open-weight models from US-based laboratories at launch. On SWE-bench Verified, which evaluates an AI's ability to resolve real GitHub software-engineering issues, Inkling scored 77.6 per cent — ahead of Nvidia Nemotron's 71.9 per cent on the same benchmark. On MMMU Pro, a multimodal reasoning benchmark, it achieved 73.5 per cent. On VoiceBench, which evaluates speech understanding, it scored 91.4 per cent, compared to Gemini 3.1 Pro's 94.4 per cent on high reasoning effort. On the Artificial Analysis Intelligence Index, a composite benchmark spanning reasoning, coding, mathematics, and general knowledge, Inkling debuted at a score of 41 — the highest published result among open-weight models from a US laboratory at time of release. These results do not place Inkling at the very front of the overall frontier, where closed models like Claude Fable 5 and GPT-5.6 Sol score considerably higher, but they establish a new standard for what is available under an open commercial licence.

The Tinker Platform: Why Inkling Is Built to Be Modified

The centre of Thinking Machines' commercial model is not Inkling itself but the Tinker fine-tuning platform, through which organisations take the open weights and customise them for specific domains and tasks. Murati's thesis is that the future of enterprise AI is not a single universal model but a landscape of differentiated, fine-tuned variants — models that know a company's codebase, data, regulatory environment, and product vocabulary. Tinker provides the infrastructure for that approach: supervised fine-tuning, preference optimisation, and evaluation tooling built specifically for Inkling's architecture. Thinking Machines has stated that Tinker, not Inkling itself, is the company's primary revenue vehicle. For organisations that have labelled training data but lack the compute to pre-train from scratch, Tinker and Inkling together offer a practical path to a domain-specific model without building from the ground up.

Availability and Hosting Partners

Full weights for the flagship model and an NVFP4 checkpoint are on Hugging Face under Apache 2.0, with no commercial restrictions on use, modification, or redistribution. Hosted inference access runs through a set of integration partners including Together AI, Fireworks, Modal, Databricks, and Baseten. Inkling-Small remains in preview at launch. The Apache 2.0 licence makes Inkling one of the most permissively licensed frontier-scale models available to date — developers and enterprises can use it commercially, fine-tune it, and build on it without royalty obligations or geographic restrictions.

What Inkling Means for Indian Software Teams

For Indian software and product teams, Inkling's release opens two practical opportunities. The first is private deployment. The open weights and Apache 2.0 licence make Inkling viable for self-hosted deployment on a team's own GPU infrastructure — in Indian data centres, on AWS Mumbai or Azure Pune, or on on-premise clusters. For teams serving banking, healthcare, or government clients with data residency requirements, a 41-billion-active-parameter multimodal model running on-premises is a meaningful alternative to routing sensitive data through US-based API endpoints.

The second is domain fine-tuning via Tinker. Indian enterprise software faces a persistent challenge: general AI models perform inconsistently on domain-specific workflows in insurance, lending, logistics, legal services, and healthcare — partly due to limited Indian-language training data, partly due to domain vocabulary gaps. Inkling as a base model, fine-tuned through Tinker on proprietary business data, reduces the barrier to a domain-adapted model. The capital cost of labelling a fine-tuning dataset and running fine-tuning steps on rented GPUs is substantially lower than pre-training a foundation model — and that delta is even more favourable in India's engineering labour market than in Western markets.

The Bottom Line

Thinking Machines Lab released Inkling on 15 July 2026: a 975-billion-parameter sparse MoE open-weight model with 41 billion active parameters, trained on 45 trillion tokens of text, image, audio, and video data. It scores 77.6 per cent on SWE-bench Verified — the top result among US open-weight models at launch — and 91.4 per cent on VoiceBench. Full weights are on Hugging Face under Apache 2.0. The Tinker platform provides fine-tuning infrastructure for teams building domain-specific models on top of Inkling. For Indian engineering teams, the combination of permissive licensing, open weights, native multimodal reasoning, and a fine-tuning platform makes Inkling a practical foundation for building privately deployed, domain-adapted AI without the cost of training from scratch.