Inkling-Small

Inkling-Small

Reads text, images and audio and replies in text, for developers building assistants, coding tools and agents. A sparse mixture-of-experts model with 276B parameters in total and 12B active per token.

Made by thinkingmachines on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Inkling-Small is an open-weights multimodal model from Thinking Machines. It takes text, images and audio as input and writes its answers as text, in English and other languages and in several programming languages.

It is a sparse mixture-of-experts transformer with 42 layers. Of its 276B parameters, 12B are active for each token.

Who it suits: Developers who build AI applications such as coding assistants, chatbots, agents that use tools and retrieval-augmented generation systems, and teams that fine-tune open models.

Industry | Who uses it | How they use it | Example
Software | Developers of coding assistants | Run a general model for code and chat | Answer questions about a code file
Product teams | Builders of agents and chatbots | Use tool calling and instruction following | Run a support chatbot that reads screenshots
Data teams | Builders of RAG systems | Answer questions over retrieved text | Combine search results with a prompt
Research | Fine-tuning researchers | Adapt open weights to a domain | Fine-tune on in-house data

Skip it if you have no server-class hardware, since the weights are about 532 GB in BF16 and about 133 GB at 4-bit by estimate.

Where to use it:

• Tinker Playground, a hosted page to try the model
• The Tinker Cookbook, which the card points to for API access
• Local serving with SGLang, vLLM, TokenSpeed, Unsloth or Hugging Face, each with its own recipe
• Third-party inference providers, which the card says also offer API access
• Windows, macOS or Linux with Python, to download the files

Use case | Who | What they get back
Coding help | Developers | Text answers and code in several programming languages
Image questions | App builders | Text answers about a picture you supply
Audio understanding | Product teams | Text output based on a WAV recording

How to set it up:

• Python and pip to run the download commands
• Server-class GPU memory for local serving, or an account with a provider
• One of the serving libraries the card lists, if you run it yourself

1. Install the Hugging Face tools.
2. Download the model files, or pick the NVFP4 build for a smaller footprint.
3. Choose a serving library from the card and follow its recipe.
4. Or skip local hosting and try the Playground or an inference provider.

pip install -U huggingface_hub
hf download thinkingmachines/Inkling-Small

How to use it: Send text, images or audio to the served model and read back text. The card gives size and length limits for images and audio that give the best results.

Input | Format and limits
Text | UTF-8 text
Image | Any pixel-based image, with each side between 40 and 4096 pixels for best results
Audio | WAV at 16 kHz, ideally under 2 minutes

Example: You serve the model with one of the listed libraries and send a photo of a chart together with a question about it. The model replies in text with its reading of the chart.

For newcomers and experts: If you are new to large models, try the hosted Playground first to see how it handles text and images, before planning any local setup. Local serving needs a lot of GPU memory.

If you know the field, a router sends each token to 6 of 256 experts, and 2 shared experts run on every token. Attention mixes local and global layers, images go through a hierarchical patch encoder, and audio is encoded as discrete tokens, all sharing one hidden space.

Text, image or audio input
Modality encoders feed one shared hidden space
Sparse MoE decoder routes each token
Text output

Good to know:

• The model type is a multimodal autoregressive transformer, and it generates text only
• The numerics supported are BF16 and NVFP4
• Hugging Face lists 54 quantized versions made by other people
• The shown part of the card has a benchmark table against open and closed weights models, but its numbers are cut off here

Title: Parameters in billions
Total parameters: 276 billion
Active per token: 12 billion

• License: Apache 2.0, and the card also links an acceptable use policy
• Cost to run: none named for the weights, though the Playground and provider APIs are separate services and need your own server-class hardware if you self-host
• Data it reaches: the text, images and audio you send it
• Not for: output other than text, since it generates text only

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 266B
• Weights: about 532 GB in BF16; about 133 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 1,114,076
• Likes: 456
• License: apache-2.0
• Quantized versions: 54
• Fine-tunes: 7
• Tags: transformers, safetensors, inkling_mm_model, image-text-to-text, conversational, audio-text-to-text

Links

• Model: https://huggingface.co/thinkingmachines/Inkling-Small
• Files: https://huggingface.co/thinkingmachines/Inkling-Small/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download thinkingmachines/Inkling-Small

Then follow the usage in its model card. Nothing in this kit runs by itself.