React Native 向け高速・特化型モデル「Apex」の紹介

Today we are announcing Apex, a React Native coding model built by Callstack.

Apex is trained for the engineering work we do every day: building React Native apps, analyzing architecture decisions, fixing framework-specific issues, and reasoning about the constraints that appear when one codebase has to ship across platforms.

The practical value is simple: React Native-specific answers, faster output, and lower inference cost.

It is already running internally with our engineers. We are now opening a private beta for selected teams while we finish the legal and operational work needed to offer it as a broader public service.

The Shift Toward Specialized Models

The broader AI coding market is moving away from a subsidization phase and toward sustainable compute economics. GitHub’s move to usage-based billing for Copilot is a clear signal: running multi-step, agentic workflows on large, general-purpose frontier models is expensive.

At the same time, smaller, heavily optimized models are proving they can alter the cost-performance curve. Public proof points like Cursor's Composer 2 or Windsurf with SWE-1 demonstrate that it is far easier to optimize a smaller model narrowly in one direction than to make a general model perfect for every workflow.

We do not believe every isolated task needs a custom model. However, repeated, high-volume engineering work, like React Native development, benefits immensely from a domain-specific model.

Why We Built Apex

React Native development is composed of multiple moving parts at once: third-party libraries, native modules, and rigid platform constraints.

General coding models know some of this, and they are getting better quickly. But broad progress does not always translate into better React Native output. A model can improve on general coding benchmarks and still miss the framework conventions, library behavior, and cross-platform details that decide whether a React Native answer is useful.

For React Native workflows and agentic engineering, we want a model that starts closer to the right answer. We need better defaults around the libraries, APIs, and patterns without needing as much tool calling, search, skill loading, or prompt scaffolding.

We built Apex to provide a focused specialist that understands these nuances natively.

What Apex Is Good For Today

We are being realistic about where it stands. Apex is not a replacement for every frontier coding model. It is a focused React Native model that is already useful and improving quickly.

To measure its capabilities objectively, we evaluate Apex against React Native Evals.

Within its specific domain, this optimized model alters the performance-to-cost ratio significantly when compared to larger, more expensive general models.

How We Trained It

Apex is based on Gemma 4, then adapted specifically for React Native coding work.

We started with proof-of-concept experiments on other bases, including Devstral and Qwen. After early testing, we moved to Gemma 4 because the base model was already stronger for React Native work before specialization.

From there, we trained Apex using Supervised Fine-Tuning (SFT) and Group Relative Policy Optimization (GRPO).

The training data comes from recent GitHub repositories and React Native ecosystem work. We did not run a random, broad web scrape; instead, the data was carefully cherry-picked around the libraries and frameworks our engineers see in daily delivery.

Our first experiments started on February 13. In March, we gathered data and trained multiple variants. Internal testing with our open-source developers started on April 2. Since then, we prepared about 50 different model variants and configurations while building the infrastructure for production-grade serving.

Where It Runs and Why It Is Fast

Speed comes from the model architecture, our serving setup, and narrow task focus.

Gemma 4 gives Apex an efficient base. We run the hosted version on a dedicated cluster of 4 to 8 NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs, depending on load. In production-serving tests, total generation output sits between 2,000 to 4,000+ tokens per second.

Because the model carries more React Native knowledge directly in its weights, it spends less time searching, calling tools, reading generic context, or relying on external skill execution before producing an answer. For developers, the practical benefit is less waiting between prompts, retries, and follow-up tasks.

We host Apex on Vast.ai Secure Cloud, using verified, trusted data-center providers. This infrastructure approach gives us the cost flexibility of rented GPU hardware combined with the strict security posture required for production enterprise serving.

Where We Go From Here

Apex is ready for more real-world use, but we are not opening it as a full public service yet.

We are still working through the legal, operational, and commercial details that come with offering a model externally. Until that is ready, we are running a private beta with selected teams so we can curate the experience, gather better feedback, and keep improving the model against real React Native work.

If you want to test Apex on React Native work, apply for the private beta below.

Table of contents

Try Apex on your React Native app

Apex is open to selected React Native teams that want to test a focused coding model against production engineering workflows.

Request access