Thinking Machines has released Inkling on Hugging Face, giving developers access to a 975-billion-parameter multimodal model designed to process text, images and audio within a claimed one-million-token context window. The release matters because it pairs unusually large scale with open availability, although running the principal checkpoints requires substantial computing capacity.

Hugging Face describes Inkling as a decoder-only Mixture-of-Experts model with 975 billion total parameters and 41 billion active during processing. According to the article, the model was trained on 45 trillion tokens spanning text, images, audio and video. Those training and context-length figures have not been independently verified in the supplied material.

The architecture uses 256 experts, selecting six routed experts while keeping two shared experts active. It combines global and sliding-window attention in a five-to-one pattern, relative positional attention and a short one-dimensional convolution intended to improve local processing. Images pass through a hierarchical patching module, while audio is converted into discrete mel-spectrogram representations.

The release includes a BF16 checkpoint and an NVFP4 version described as calibrated for lower-precision inference. Hugging Face reports day-one support in Transformers, SGLang, vLLM and llama.cpp, along with speculative multi-token prediction layers intended to accelerate generation. Developers can also access the model through hosted inference providers rather than operating the hardware themselves.

Inkling Expands Multimodal Model Access

Inkling gives researchers and developers a new option for applications that must reason across language, sound and visual material without relying on separate models for every input type. Thinking Machines is positioning it for domain-specific fine-tuning, which could make it relevant to scientific, media-analysis and enterprise workflows built around mixed data.

Its scale also limits who can operate it directly. Hugging Face says the BF16 checkpoint requires about 2 TB of VRAM, while the NVFP4 checkpoint needs roughly 600 GB of VRAM. Those requirements place full deployment beyond typical consumer systems and many smaller research teams, making hosted access or heavier quantization the more practical route for many users.

AI Systems Performance Engineering: Optimizing Model Training and Inference Workloads with GPUs, CUDA, and PyTorch

View Latest Price

As an affiliate, we earn on qualifying purchases.

As an affiliate, we earn on qualifying purchases.

Inside Inkling’s Sparse Architecture

Inkling belongs to a growing group of sparse Mixture-of-Experts models that contain many parameters but activate only a fraction for each input. Its 41-billion active-parameter count is intended to reduce inference work compared with activating all 975 billion parameters, though real-world speed and cost depend on implementation and hardware.

The model’s multimodal towers are comparatively simple: image patches and audio chunks are converted into embeddings that the main decoder can process alongside text. Image inputs also carry a temporal dimension for possible video processing, but Hugging Face said out-of-the-box video performance was not evaluated. The article focuses on architecture and deployment examples rather than comparative benchmark tables.

“This model is huge.”

— Hugging Face’s Inkling release article

Benchmarks and Licensing Still Missing

The supplied announcement does not include independent benchmark results, detailed safety evaluations or direct comparisons with other large multimodal models. It is also unclear how reliably the one-million-token context window performs across text, audio and image-heavy workloads, or how much its speculative prediction layers improve speed in production.

The material calls Inkling an open model but does not specify the license, usage restrictions or whether training code and complete data documentation are available. Pricing for hosted inference, fine-tuning requirements and performance on consumer-grade quantized deployments are also not provided.

Independent Testing Moves to Developers

Developers are expected to begin testing Inkling through Transformers and supported inference engines, while organizations with suitable hardware can compare the BF16 and NVFP4 checkpoints. Early evaluations should clarify latency, memory use and multimodal accuracy under practical workloads.

Attention will also turn to model cards, licensing terms and benchmark disclosures from Thinking Machines or Hugging Face. Video evaluations, third-party safety testing and fine-tuned domain models would provide a clearer measure of where Inkling performs well and where its limitations remain.

Key Questions

What is Inkling?

Inkling is a multimodal Mixture-of-Experts model from Thinking Machines that accepts text, images and audio. Hugging Face lists 975 billion total parameters and 41 billion active parameters.

Can Inkling process video?

The architecture accepts image inputs with a temporal dimension that may support video-related fine-tuning. Hugging Face says native video performance has not been evaluated, so no firm capability claim can yet be made.

Can Inkling run on a personal computer?

The primary checkpoints are unlikely to fit on an ordinary computer. Hugging Face estimates 2 TB of VRAM for BF16 and about 600 GB for NVFP4, though llama.cpp-compatible quantized versions may offer other deployment options.

Which software supports Inkling?

The announcement reports day-one support in Transformers, SGLang, vLLM and llama.cpp. It also describes access through serverless inference providers.

Has Inkling’s performance been independently verified?

No independent evaluation is included in the supplied material. Claims about its context capacity, reasoning ability and inference performance should remain attributed to Thinking Machines and Hugging Face until third-party testing is available.

Source: Hugging Face