Uploaded February 2026 | Updated September 2026, 8 hours ago
In this episode of React Native AI Unpacked, Artur introduces @react-native-ai/mlc and explains how MLC-LLM runs on-device inference across iOS and Android.
We cover how MLC optimizes inference for the models you choose, including techniques like memory planning, operator fusion, hardware-specific optimizations, and library offloading. The goal is vendor-agnostic acceleration that can use GPU drivers like Vulkan or OpenCL when available, and fall back to CPU optimizations when it is not.
Then we walk through the demo app from the repository: selecting a model, downloading it, preparing it, and using three interaction modes. You will see generateText, streamText, and generateObject (structured output) via the Vercel AI SDK, with notes on parsing errors and model limits.
Links:
- react-native-ai repository: github.com/callstackincubator/ai
- MLC-LLM: llm.mlc.ai
If you run into issues, check the docs in the repo or open an issue on GitHub. Star the project if it helps.
## Chapters
00:00 Intro and what this episode covers
00:20 What MLC-LLM is and where to find it
00:29 Model choice and build-time optimizations
00:43 Optimization techniques (memory planning, fusion, offloading)
00:56 Hardware acceleration across devices
01:02 GPUs & drivers (OpenCL, Vulkan)
01:20 CPU optimizations on low-end devices
01:31 Platform and backend support overview
01:55 Why third-party models matter (flexibility)
02:03 Bring your own model and Hugging Face
02:14 AI SDK support and consistent API
02:45 Demo app overview
03:06 Model list and Hugging Face config
03:47 What the library API covers
04:07 Download and prepare the selected model
04:25 Interaction modes overview
04:28 generateText
04:38 streamText
04:47 generateObject and parsing errors
05:22 Running generateText in the demo
05:50 generateObject with Zod schema
07:10 streamText and token buffer
07:54 Integration recap and performance
In this episode of React Native AI Unpacked, Artur introduces @react-native-ai/mlc and explains how MLC-LLM runs on-device inference across iOS and Android.
We cover how MLC optimizes inference for the models you choose, including techniques like memory planning, operator fusion, hardware-specific optimizations, and library offloading. The goal is vendor-agnostic acceleration that can use GPU drivers like Vulkan or OpenCL when available, and fall back to CPU optimizations when it is not.
Then we walk through the demo app from the repository: selecting a model, downloading it, preparing it, and using three interaction modes. You will see generateText, streamText, and generateObject (structured output) via the Vercel AI SDK, with notes on parsing errors and model limits.
Links:
- react-native-ai repository: github.com/callstackincubator/ai
- MLC-LLM: llm.mlc.ai
If you run into issues, check the docs in the repo or open an issue on GitHub. Star the project if it helps.
## Chapters
00:00 Intro and what this episode covers
00:20 What MLC-LLM is and where to find it
00:29 Model choice and build-time optimizations
00:43 Optimization techniques (memory planning, fusion, offloading)
00:56 Hardware acceleration across devices
01:02 GPUs & drivers (OpenCL, Vulkan)
01:20 CPU optimizations on low-end devices
01:31 Platform and backend support overview
01:55 Why third-party models matter (flexibility)
02:03 Bring your own model and Hugging Face
02:14 AI SDK support and consistent API
02:45 Demo app overview
03:06 Model list and Hugging Face config
03:47 What the library API covers
04:07 Download and prepare the selected model
04:25 Interaction modes overview
04:28 generateText
04:38 streamText
04:47 generateObject and parsing errors
05:22 Running generateText in the demo
05:50 generateObject with Zod schema
07:10 streamText and token buffer
07:54 Integration recap and performance
