Multimodal input support
Inkling is a multimodal autoregressive transformer that accepts text, images, and audio, while producing text outputs. The model card says it is intended for English and other languages, as well as multiple coding languages.
Inkling
소유권 인증Inkling is Thinking Machines Lab’s open-weights multimodal foundation model for text, image, and audio inputs, with fine-tuning on Tinker.
Overview
Inkling is Thinking Machines Lab’s open-weights multimodal model. It is presented as a general-purpose foundation model that accepts text, image, and audio inputs and generates text outputs, with support for English, other languages, and multiple coding languages.
The model card describes Inkling as a 66-layer decoder-only transformer with sparse Mixture-of-Experts routing, 975B total parameters, 41B active, and a context window of up to 1M tokens. Thinking Machines positions it as a customizable base model for developers building conversational systems, agentic workflows, coding tools, and retrieval-augmented applications.
Inkling is distributed with open weights under Apache 2.0, available through Hugging Face, and can be fine-tuned on Tinker. The launch materials also point to Inkling-Small as a lighter-weight preview in the same family, but the source page emphasizes Inkling as the main release.
Features
Sparse MoE architecture
The model uses a sparse Mixture-of-Experts architecture with 975B total parameters and 41B active parameters. According to the model card, each token is routed to 6 of 256 experts, plus 2 shared experts.
Long-context processing
Inkling supports a context window of up to 1M tokens, which makes it suitable for long-document and long-conversation workflows.
Controllable reasoning effort
The launch post says Inkling balances cost and performance through controllable thinking effort. That makes it a base model that can be used with different levels of reasoning effort depending on the task.
Fine-tuning via Tinker
Inkling is available for fine-tuning on Tinker, and the Tinker product page describes training through four functions: forward_backward, optim_step, sample, and save_state.
Multiple access paths
The model card says the weights are available through Hugging Face, and the model is also accessible via Tinker and third-party inference providers.
Use cases
Multimodal assistants
Developers can build chatbots or assistant-style products that need to respond to text prompts and also interpret images or audio in the same session.
Developer applications
Teams working on coding copilots, tool-using agents, or retrieval-augmented systems can use Inkling as a general-purpose base model for downstream tuning and integration.
Model customization
Researchers and product teams can fine-tune the open weights on Tinker to adapt the model to a specific behavior or workflow, such as a custom response style or domain task.
Long-context workflows
Applications that must process long documents, lengthy conversations, or extended context across multiple turns can take advantage of the model’s 1M-token context window.
Deployment planning
Organizations evaluating open-weights deployment options can use the model card’s checkpoint and hardware guidance to plan inference on their own systems or through supported providers.
Pros and Cons
Pros
- Supports text, image, and audio inputs in a single general-purpose model.
- Open weights and Apache 2.0 licensing make the model available for research and downstream customization.
- Long context support up to 1M tokens broadens its use for extended documents and agent workflows.
- Tinker provides a documented fine-tuning path, including a basic API workflow for training and resumption.
- The model card gives concrete deployment guidance, including checkpoint formats and hardware requirements.
Cons
- The source materials do not include pricing details, so the commercial structure is unclear from the public pages provided.
- Running the BF16 checkpoint locally requires substantial hardware; the model card says at least 2 TB of aggregated VRAM is needed for that format.
FAQ
How is Inkling distributed?
Inkling is available as an open-weights model, with the weights offered through Hugging Face and the model also accessible via Tinker and third-party inference providers. The announcement and model card indicate it is intended for downstream developers rather than a consumer chat app.
What can Inkling take as input and output?
The model card says Inkling accepts text, image, and audio inputs and generates text outputs. It is designed for general-purpose conversational use, instruction-following, coding, agentic workflows, and retrieval-augmented generation systems.
How does fine-tuning work with Inkling?
Tinker is the training API used to fine-tune Inkling. The model card says Inkling is available for fine-tuning on Tinker, and the launch post describes using Tinker to write, run, and evaluate a self-finetuning job before updating the active weights.
What are Inkling's core model characteristics?
The model card describes Inkling as a 66-layer decoder-only transformer with sparse Mixture-of-Experts routing. It has 975B total parameters, 41B active, and supports a context window of up to 1M tokens.
Is there public pricing information for Inkling or Tinker?
The source materials do not provide pricing information. The pricing URL currently returns a 404 page not found.
Quick Facts
- Category
- Open-weights multimodal model
- Provider
- Thinking Machines Lab, Inc.
- Release date
- July 15, 2026
- License
- Apache 2.0
- Inputs
- Text, image, audio
- Outputs
- Text
Inkling 대안
Orca
Orca is an Agent Development Environment for shipping with coding agents, running multiple CLI agents in parallel across isolated worktrees, with desktop and mobile workflows.
Firebase Studio
Firebase Studio is a web-based workspace for full-stack app development with Gemini-assisted coding, app previews, cloud emulators, collaboration, and browser deployment.
OpenAI
OpenAI offers ChatGPT, API, Platform tools, and Codex for conversational AI, model building, and research and product updates.
AakarDev AI
AakarDev AI helps teams manage AI provider access, project setups, logs, and analytics in one dashboard with BYOK support.
JetBrains
JetBrains makes developer tools for individuals and teams, including IntelliJ IDEA, PyCharm, DataGrip, WebStorm, Rider, and CLion, with built-in run/debug/test workflows.
书生
书生 is a general-purpose large model suite for language, multimodal, weather, ocean, industrial design, 3D, finance, and scientific discovery tasks.