Inkling

Inkling

Reads text, images and audio and writes text, as a huge open multimodal model for coding assistants, agents and chatbots. For developers with multi-GPU servers or a hosted API.

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 is a general-purpose multimodal model from Thinking Machines. It accepts text, images and audio and writes text back, in English and other languages and across coding languages. The card lists 975B parameters in total with 41B active, built as a sparse mixture of experts.

Who it suits: It suits developers who build coding assistants, agents with tools and chatbots, and who have the hardware or an API route for a model this large.

Industry | Who uses it | How they use it | Example
Software development | Developers | Build coding assistants and tool-using agents | Running an assistant that calls outside tools
Customer support | Product teams | Build chatbots that read text, pictures and audio | A chatbot that answers questions about a screenshot
Media and meetings | Audio and video teams | Turn recordings into written text | Sending a WAV recording and asking for a summary
AI research | Model researchers | Compare large open-weight multimodal models | Reading the card's benchmark table

You can skip it if you have a single consumer GPU, since the BF16 weights alone come to about 1,905 GB.

Where to use it:

• SGLang, vLLM, TokenSpeed, Unsloth and Hugging Face, each with a recipe page the card links
• The Tinker Playground and Tinker Cookbook, for trying it or calling it through an API
• 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 | Written code and explanations
Image and audio questions | Product teams | Text answers about a picture or a WAV recording
Agent work | Agent builders | Text output that drives tool calls

How to set it up: You need Python for the download tool and a multi-GPU server. The BF16 files are about 1,905 GB, or about 476 GB at 4-bit as an estimate. The card also links an NVFP4 version named Inkling-NVFP4.

• Python on Windows, macOS or Linux
• The huggingface_hub package
• One of the runtimes listed above
• A server with hundreds of gigabytes of GPU memory

1. Install huggingface_hub.
2. Download the model files, or pick the NVFP4 version.
3. Follow the recipe for your runtime to serve it.

pip install -U huggingface_hub
hf download thinkingmachines/Inkling

How to use it: The visible card gives no run command, only recipe pages for each runtime. Once the model is served, you send text, an image or audio and read the text it writes back.

Input | What it accepts
Text | UTF-8 text
Image | Any pixel-based format, each side ideally from 40 to 4096 pixels
Audio | WAV at 16 kHz, ideally under 20 minutes

Example: You send a WAV recording of a meeting and ask for a summary. The model reads the audio and writes the summary as text.

For newcomers and experts: If you are new to this, it is a huge assistant that can read, look and listen, then answer in writing. A first small try is the hosted playground, since running it yourself needs a big server.

For people who know the field, it stacks 66 decoder layers. A router sends each token to 6 of 256 experts, and 2 shared experts see every token. Attention mixes local and global layers. Images and video go through a hierarchical patch encoder, and audio is encoded as discrete tokens.

Title: HLE text-only scores in the card's table
Inkling: 29.7%
Nemotron 3 Ultra: 26.6%
Kimi K2.5: 29.4%
Kimi K2.6: 35.9%
GLM 5.2: 40.1%
DeepSeek V4 Pro: 35.9%

You send text, an image or audio
The model encodes each type of input
A sparse expert network processes the tokens
It writes a text answer

Good to know: The scores come from the authors, who report Inkling at effort 0.99 and dated the comparison scores July 14, 2026. The table also lists closed-weight models that rank higher on this test.

• License: Apache 2.0
• Cost to run: none named, and third-party providers set their own terms
• Data it reaches: the text, images and audio you send, plus the downloaded files
• Not for: output other than text, or machines without server-class GPU memory

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 952B
• Weights: about 1,905 GB in BF16; about 476 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 1,583,506
• Likes: 1,847
• License: apache-2.0
• Quantized versions: 47
• Fine-tunes: 16
• Tags: transformers, safetensors, inkling_mm_model, image-text-to-text, conversational, audio-text-to-text

Links

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

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download thinkingmachines/Inkling

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