Andrey Kolkov

Posted on Dec 2 • Originally published at github.com

I Skipped My Birthday to Give Go Its First Real ML Framework

#go #machinelearning #opensource #deeplearning

TL;DR: On December 1st, 2025 - my birthday - I released Born v0.5.0. Not because the version number was nice, but because that was the day we successfully trained our first Hierarchical Reasoning Model in pure Go. Nine months of preparation. Two weeks of intense labor. One framework that makes production ML in Go a reality.

The Dream

I've been writing Go for years. Love the language. The simplicity. The tooling. The "it just works" deployment story.

But every time I needed machine learning? Back to Python. TensorFlow. PyTorch. Docker images. Dependency hell. The whole circus.

For years I kept asking myself: Why can't we do ML in Go?

Sure, there were attempts. Gorgonia. GoMLX. Brave projects, but none of them felt like a complete answer. None of them made me think "yes, I can build production ML systems in pure Go."

So I decided to build one myself.

Nine Months of Obsession

Born didn't happen in a weekend hackathon. It was extracted from a private project we'd been building for nine months. Yes, nine months - like bringing a child into the world.

It started with a fundamental problem: we needed fast GPU acceleration in pure Go, without CGO. So we built goffi - a Foreign Function Interface for Go. Then we created Go bindings for WebGPU, making GPU compute accessible without a C compiler.

Layer by layer, we assembled the foundation:

Tensor operations with automatic broadcasting
Automatic differentiation engine
Neural network modules (Linear, Conv2d, Attention, Transformers)
Optimizers (SGD, Adam)
GPU acceleration via WebGPU

All of this was running in our internal systems, powering real workloads. But it wasn't ready for the world yet.

The Birth

Two weeks before my birthday, we decided: it's time to open source this.

We extracted the ML framework from our monorepo, cleaned up the API, wrote documentation, and pushed to GitHub. But a repo without polish isn't a product.

So for two weeks, I barely slept:

Implemented complete transformer architecture support
Added attention mechanisms (SDPA, Multi-Head, Grouped Query Attention)
Built KV-cache for efficient LLM inference (3.94x speedup!)
Created tokenizers and text generation pipeline
Fixed dozens of edge cases
Wrote hundreds of tests

Then came December 1st.

My birthday. And also the day our internal HRM model - a Hierarchical Reasoning Model - successfully completed its first training run on Born. This architecture, which outperforms o3-mini and Claude 3.7 on reasoning benchmarks with just 27M parameters, was now trainable in pure Go.

That's when I knew: Born was ready.

I didn't celebrate with cake. I celebrated by pushing v0.5.0 to GitHub and watching the first stars appear.

What is Born?

Born is a modern deep learning framework for Go, inspired by Burn - the excellent Rust ML framework.

The philosophy is simple: "Models are born production-ready."

import "github.com/born-ml/born"

model := born.Load("model.born")
prediction := model.Predict(image)

// That's it. No Python. No containers. Just Go.

What you get:

Pure Go - No CGO, trivial cross-compilation
GPU Acceleration - WebGPU backend with 123x speedup
Type Safety - Generics-powered API with compile-time guarantees
Single Binary - Deploy your model as one executable
Modern Architectures - Transformers, Attention, RMSNorm, SwiGLU, GQA

Why Does This Matter?

Because Go is everywhere in production.

Kubernetes. Docker. Terraform. Prometheus. The entire cloud-native ecosystem runs on Go. And now, when you need to add ML inference to your Go microservice, you don't need to:

Spin up a Python sidecar
Manage gRPC/REST bridges
Deal with model serving infrastructure
Explain to your ops team why you need CUDA drivers

You just import Born and go.

The Numbers

What	Result
Development	9 months + 2 weeks
GPU Speedup	123x (MatMul 1024x1024)
KV-Cache Speedup	3.94x
MNIST Accuracy	98.18% (CNN)
Dependencies	Zero (pure Go)
Tests	400+
Deployment	Single binary

We Need You

Born is live. It works. But it's just getting started.

We're looking for:

Testers with real projects - Try Born in your production systems. Find the edge cases. Push the limits. Your feedback shapes the roadmap.
Contributors - Whether it's a bug fix, new operator, documentation improvement, or a whole new backend - we welcome it all.
Early adopters - Be the first to deploy ML in pure Go. Share your experience. Help grow the community.

The best frameworks aren't built in isolation. They're shaped by real-world usage.

What's Next?

The roadmap includes:

ONNX Import - Train in PyTorch, deploy in Go
Quantization - INT8/FP16 for smaller, faster models
More Backends - Vulkan, Metal, maybe CUDA
Flash Attention - Even faster transformers

But honestly? The core is done. You can use Born today to build real ML systems in Go.

Try It

go get github.com/born-ml/born@v0.5.3

Check out the GitHub repo for examples and documentation.

Star the repo if you believe Go deserves great ML tooling.

December 1st, 2025 was my birthday. But it was also the day Go's machine learning framework was born.

Nine months of preparation. Two weeks of intense labor. One framework that changes everything.

I couldn't have asked for a better birthday present.

Now let's build something amazing together.

DEV Community