# Thinc: Functional Deep Learning Model Composition from Explosion AI > Thinc is a Python deep learning library from the makers of spaCy that builds neural networks through function composition rather than class inheritance, with integrated type checking and wrappers for PyTorch, TensorFlow, and MXNet. It suits engineers who want fine-grained control over model structure and a config system that separates hyperparameters from code. **explosion/thinc** — 🔮 A refreshing functional take on deep learning, compatible with your favorite libraries - Repository: https://github.com/explosion/thinc - Website: https://thinc.ai - Stars: 2,890 · Forks: 293 - Language: Python - License: MIT - Published: 2026-09-10 · Updated: 2026-09-10 · Language: en - Canonical page: https://hysenlabs.com/projects/explosion-thinc ## Type-Checked Composition Over Class Hierarchies Most deep learning frameworks build models through class inheritance: subclass a Module or Layer, override a forward method, and compose via container classes. Thinc takes a different path. Models are assembled by composing functions, each carrying custom type annotations that a mypy plugin can verify at static analysis time. A model in Thinc is a typed object wrapping two functions: one that produces an output from an input (the forward pass) and one that receives an upstream gradient and returns a downstream gradient (the backward pass). Because both directions are explicit in the type signature, composition errors can surface before a training run starts rather than at runtime. The project describes itself as a "refreshing functional take on deep learning": there is no global graph, no session, and no symbolic compilation step. This reduces certain classes of bugs but places more cognitive demand on the developer, who must understand what types are flowing through each layer rather than relying on a framework to infer shapes automatically. ## Installing Thinc and One Environment Caveat Thinc runs on Linux, macOS, and Windows. The README recommends keeping pip, setuptools, and wheel up to date before installing: ```bash pip install -U pip setuptools wheel pip install thinc ``` The README includes one specific warning for Python 3.7 and later: if PyTorch is already installed, it may have added a `dataclasses` package that conflicts with the standard library. Remove it before installing Thinc: ```bash pip uninstall dataclasses ``` The base package does not pull in PyTorch, TensorFlow, or MXNet. Those must be installed separately when you need the framework wrappers. For GPU support and optional backend dependencies, the extended installation documentation at thinc.ai/docs/install covers the additional steps. ## The Config System and Function Registry Thinc includes a config system designed to describe trees of objects and hyperparameters in plain configuration files. The registry mechanism lets you register custom Python functions under a name, then reference those names in a config file. This decouples model architecture from training scripts: the same config file can drive different training pipelines without code changes, and hyperparameter search tools can modify the config rather than the source. The README notes that the config system is the same one spaCy uses to manage its component registry, which means engineers already familiar with spaCy's config format will find Thinc's version familiar. The tradeoff is indirection: tracing a model definition from a config file back to the Python function that implements it requires knowing how the registry lookup works. The full documentation is at thinc.ai/docs/usage-config. ## Wrapping PyTorch, TensorFlow, and MXNet Layers Thinc's framework integration lets you place a PyTorch module, a TensorFlow layer, or an MXNet block inside a Thinc model and treat it as a composable unit. The README describes this at thinc.ai/docs/usage-frameworks. One notebook in the examples directory demonstrates the practical pattern: a part-of-speech tagger that combines a BERT model from the Hugging Face transformers library (running on PyTorch) with a custom output layer written directly in Thinc, configured from a single file. The interoperability does not abstract the underlying backends away entirely. You still need the target framework installed, and gradients flow through each framework's own autograd system. The benefit is mixing: you are not forced to pick a single backend for an entire model when different layers come from different sources. ## What Thinc Does Not Provide Thinc is not an end-to-end training framework. The README and documentation do not describe a built-in training loop, checkpointing system, or distributed training facility in the core library. Teams expecting the batteries-included experience of Keras or PyTorch Lightning will need to build these pieces themselves or follow the example notebooks. The examples directory includes a parallel training notebook using Ray, but the README presents this as an additional example rather than a first-class feature. The functional API is also a genuine learning investment: engineers accustomed to subclassing torch.nn.Module will find the composition model unfamiliar until they work through the introductory notebook. PyTorch Lightning, by comparison, provides a structured training loop, logging hooks, and multi-GPU support out of the box, at the cost of an opinionated class-based structure that Thinc avoids. Static type checking via the mypy plugin is an optional step. The README points to a separate setup guide at thinc.ai/docs/install#type-checking. Teams that skip type checking lose one of Thinc's stronger design properties and reduce the advantage over less opinionated alternatives. ## Maintenance Cost and License The last push to the Thinc repository was on 2026-03-27, with release v8.3.13 published on 2026-03-23. The repository is not archived. Explosion AI, the team behind spaCy and Prodigy, maintains Thinc, so its long-term maintenance is tied to those products continuing to use it as their model layer. The MIT license places no restrictions on commercial use and permits modification and redistribution. Upgrade pressure tends to arrive from spaCy version changes rather than from Thinc's own release schedule. The pyproject.toml shows Cython, Pydantic v2, and NumPy as compile-time or runtime requirements; a major version change in any of those will require attention. The requirements.txt pins `confection>=1.3.2,<2.0.0`, which is the config library that underpins Thinc's config system and follows its own release cadence. Release v8.3.11 added full Pydantic v2 support via the confection v1 upgrade, which means teams running older Pydantic environments must upgrade when moving to the current Thinc line. The binary wheel build matrix in pyproject.toml skips Python 3.8 and 3.9, reflecting the project's minimum Python version requirements. ## Conclusion Thinc is worth adopting for engineers already in the spaCy or Prodigy ecosystem, or for anyone building custom model architectures that mix framework layers. Verify whether your team can accept the functional composition model before committing: engineers familiar only with class-based frameworks like Keras will need a ramp-up period. The MIT license has no restrictions on commercial use. The last push to the repository was on 2026-03-27, and because Thinc is a core dependency of spaCy, its upgrade schedule is tied to spaCy releases rather than standing alone. ## FAQ ### Can Thinc wrap models from different frameworks in the same pipeline? Yes. The README describes wrapping PyTorch, TensorFlow, and MXNet models for use within a single Thinc network. Each wrapped layer behaves as a composable Thinc model object that fits into the same functional composition chain. ### Is Thinc the model layer used by spaCy? The README states that previous versions of Thinc have been running in production via both spaCy and Prodigy, and that the current version was written specifically to let users compose, configure, and deploy custom models built with their preferred framework. ### Does Thinc require a GPU to run? The base pip install does not require a GPU. The README notes that the extended installation documentation at thinc.ai/docs/install covers optional dependencies for different backends and GPU support. ## Sources - [explosion/thinc on GitHub](https://github.com/explosion/thinc) - [License: MIT](https://github.com/explosion/thinc/blob/v8.3.x/LICENSE) - [Project website](https://thinc.ai) - [README](https://github.com/explosion/thinc/blob/v8.3.x/README.md) - [Releases](https://github.com/explosion/thinc/releases) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/explosion-thinc