Inkling: Thinking Machines' 1 Trillion Parameter Multimodal Open Model
AIOpen Sourceinklingthinking-machinesmultimodalllmopen-sourcemoespeculative-decodingmtp
Loading…
On 15 July 2026, Thinking Machines released Inkling, a ~1-trillion-parameter open multimodal model that natively accepts image, text, and audio inputs. It is the first large open model to combine a 1M-token context window with native any-to-any modality support — no modality-specific encoders, no projection layers, no adapter modules. The model understands audio waveforms, pixels, and text as first-class input types.
Inkling is built on a decoder-only MoE architecture trained from scratch on 45 trillion tokens of text, images, audio, and video. It uses a unified any-to-any pipeline that processes all modalities through the same transformer backbone, with modality-specific pre-processors that operate at the token embedding level rather than through separate encoders.
| Component | Specification |
|---|---|
| Total parameters | ~1T |
| Architecture | Decoder-only MoE Transformer |
| Training data | 45T tokens (text + images + audio + video) |
| Context window | 1M tokens |
| Input modalities | Image, text, audio (native) |
| Weight formats | BF16 (2 TB VRAM), NVFP4 (600 GB VRAM) |
| Speculative decoding | MTP (multi-token prediction) layers |
| Day-0 engines | transformers 5.14.0, vLLM, SGLang, llama.cpp |
Multi-Token Prediction (MTP) adds extra layers that predict several tokens at once rather than just the next one. During inference, these extra layers act as drafters for speculative decoding, speeding up generation by a measurable multiplier without compromising output quality. The MTP drafter ships as part of the release.
What distinguishes Inkling from other multimodal models is that audio and images are not secondary capabilities bolted on via separate encoders. The model's any-to-any pipeline means audio waveforms are tokenised directly and processed through the same attention mechanisms as text tokens. This unified representation enables cross-modal reasoning — understanding the relationship between spoken words, visual context, and written text within a single forward pass.
The practical implication for enterprise deployments: a single model can process a video call recording, a technical diagram, and a chat log simultaneously, relating information across modalities without separate pipeline stages.
Inkling is large by any standard. The BF16 checkpoint requires 2 TB of VRAM, while the NVFP4 variant requires 600 GB. For most enterprise teams, deployment means:
Thinking Machines provides SLURM scripts for cluster deployment alongside the standard vLLM and SGLang serving configurations.
The release includes tinker, a managed tool for post-training Inkling. The cookbook covers fine-tuning, distillation, and reinforcement learning workflows. The recommended approach for downstream task adaptation is knowledge distillation using Inkling as a teacher model, leveraging the GOLD algorithm to match token logits across different tokenizers — enabling distillation to any model on the Hub regardless of vocabulary.
Inkling changes the multimodal deployment calculus in three ways:
Single-model architecture eliminates pipeline complexity. One model processes all modalities, reducing the engineering surface area for multimodal applications.
Open weights at this scale are a strategic asset. A 1T-parameter open multimodal model lets enterprises evaluate frontier-level multimodal capability on their own infrastructure before committing to API-based consumption models.
Distillation from Inkling is a viable path to smaller deployed models. The GOLD-based distillation pipeline means enterprises can use Inkling as a teacher to train smaller, faster models for specific deployment contexts.
For teams building multimodal document processing, media analysis, or any application where understanding the relationship between different data types is critical, Inkling represents the current state of the art in open multimodal AI.