Inkling-Small

Description

Inkling-Small is a general-purpose multimodal autoregressive transformer that accepts text, image, and audio inputs and generates text output. It is intended for use across English and other languages, as well as multiple programming languages.

The model is designed for developers building agentic and tool-use systems, coding assistants, chatbots, retrieval-augmented generation systems, instruction-following applications, and other natural-language or multimodal workflows. It is released with open weights to support research, fine-tuning, and integration into third-party products.

This model is ready for commercial or non-commercial use.

Third-Party Community Consideration

This model is not owned or developed by NVIDIA. This model has been developed and built to a third-party's requirements for this application and use case; see the non-NVIDIA Thinking Machines Lab Inkling-Small Model Card.

License and Terms of Use

GOVERNING DOWNLOAD TERMS: Use of this model is governed by the NVIDIA Open Model Agreement.

ADDITIONAL INFORMATION: Apache License 2.0.

Deployment Geography

Global

Use Case

Developers can use Inkling-Small to build multimodal assistants, agentic and tool-use systems, coding assistants, chatbots, retrieval-augmented generation systems, and general conversational or instruction-following applications that accept text, image, or audio inputs and produce text.

Inkling-Small is intended for research, fine-tuning, and integration into third-party products. Downstream developers should perform use-case-specific evaluation and apply appropriate safeguards, monitoring, and human oversight before deployment, particularly in high-stakes or safety-critical applications.

Release Date

[Hugging Face] 07/27/2026 via Inkling-Small on Hugging Face. [NGC Catalog] 11/05/2026 via Inkling-Small

References

• Inkling-Small on Hugging Face
• Inkling-Small BF16 weights
• Inkling-Small NVFP4 weights
• Tinker Playground
• Tinker Cookbook
• SGLang recipe
• vLLM recipe
• Training Data Documentation

Model Architecture

Architecture Type: Transformer
Network Architecture: Multimodal autoregressive, decoder-only sparse Mixture-of-Experts (MoE) transformer
Number of Model Parameters: 276B total, 12B active
Number of Layers: 42
Numerics Support: BF16 and NVFP4

Inkling-Small uses a sparse MoE feed-forward backbone. Each token is routed to 6 of 256 experts, with 2 additional shared experts active for every token. The attention stack combines local and global layers. Images are processed using a hierarchical patch encoder, while audio is represented using discrete token encoding. All supported modalities are projected into a shared hidden space and processed jointly by the decoder.

Input

Input Type(s): Text, Image, Audio Input Format(s): UTF-8 text, Red, Green, Blue (RGB), WAV audio
Input Parameters: One-Dimensional (1D), Two-Dimensional (2D), and One-Dimensional (1D) Other Properties Related to Input: For optimal performance, each image dimension should be between 40 and 4,096 pixels. Audio should use WAV format sampled at 16 kHz and should ideally be shorter than two minutes.

Output

Output Type(s): Text
Output Format: UTF-8 encoded string
Output Parameters: One-Dimensional (1D) sequences
Other Properties Related to Output: The model produces natural-language and code output.

Our AI models are designed and/or optimized to run on NVIDIA GPU-accelerated systems. By leveraging NVIDIA's hardware (e.g. GPU cores) and software frameworks (e.g., CUDA libraries), the model achieves faster training and inference times compared to CPU-only solutions.

Software Integration:

Runtime Engines:

• SGLang
• vLLM

Supported Hardware:

• NVIDIA Blackwell B200, NVIDIA Blackwell B300, NVIDIA Blackwell GB300, NVIDIA Blackwell GB200
• NVIDIA Hopper H100, NVIDIA Hopper H200, NVIDIA Hopper H100 NVL

Preferred Operating Systems: Linux

The integration of foundation and fine-tuned models into AI systems requires additional testing using use-case-specific data to ensure safe and effective deployment. Following the V-model methodology, iterative testing and validation at both unit and system levels are essential to mitigate risks, meet technical and functional requirements, and ensure compliance with safety and ethical standards before deployment.

Model Version(s)

• Inkling-Small BF16
• Inkling-Small NVFP4

Training, Testing, and Evaluation Datasets

Training Dataset

Data Modality: Text, Image, Audio, Video
Text Training Data Size: Undisclosed
Image Training Data Size: Undisclosed
Audio Training Data Size: Undisclosed
Video Training Data Size: Undisclosed
Data Collection Method by Dataset: Hybrid: Automated, Manually-Collected, Synthetic .
Labeling Method by Dataset: Undisclosed
Properties: The data-curation process included cleaning, processing, and modifying datasets. Depending on data type, processing could include deduplication and filtering to remove junk or other low-quality content or to advance safety and other objectives.

Testing Dataset

Data Collection Method by Dataset: Undisclosed
Labeling Method by Dataset: Undisclosed
Properties: Undisclosed

Evaluation Dataset

Evaluation Benchmark Score: The Inkling-Small source model card reports results across agentic coding, general agentic performance, reasoning, factuality, chat, safety, vision, and audio. Selected reported Inkling-Small results include:

Data Collection Method by Dataset: Hybrid: Automated, Manually-Collected.
Labeling Method by Dataset: Undisclosed
Properties: Evaluation covers coding agents, tool use, general and abstract reasoning, factuality, instruction following, multilingual knowledge, safety, vision understanding, and audio understanding. See the linked third-party model card for the complete comparison tables, harness notes, and methodology. NVIDIA has not independently reproduced the linked results.

Inference

Acceleration Engine: SGLang and vLLM are supported through public deployment recipes.
Test Hardware: NNVIDIA Blackwell (GB200, B200, B300, GB300), NVIDIA Hopper (H200)

Known Limitations

Inkling-Small may exhibit limitations common to foundation models, including hallucinations, occasional failures to follow instructions precisely, and degraded performance in long, multi-turn conversations. The model may reflect demographic, cultural, or linguistic biases in its web-derived and synthetic training data and may perform unevenly across languages, dialects, or subject areas that were less represented during training.

The model's knowledge is limited to information available as of its training cutoff, which is not specified in the source model card. It may not reflect later events or developments.

Downstream developers should:

• Evaluate performance, safety, and fairness for their specific use case, language, and population before deployment.
• Apply content filtering, rate limiting, monitoring, and other application-layer safeguards.
• Use human review for high-stakes or safety-critical outputs.
• Avoid medical, legal, or other safety-critical decision-making deployments without additional fine-tuning, domain-specific validation, and human oversight.

Ethical Considerations

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. Developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.

Please make sure you have proper rights and permissions for all input image and video content; if image or video includes people, personal health information, or intellectual property, the image or video generated will not blur or maintain proportions of image subjects included.

Please report model quality, risk, security vulnerabilities or NVIDIA AI Concerns here.

Get Help

Enterprise Support

For help with NVIDIA NIM, customers with an active NVIDIA AI Enterprise subscription can file a support ticket.

NVIDIA Developer Community Forum

Access community knowledge-base articles and support cases through the NVIDIA Developer Forums.