React Native ExecuTorch v0.11.0 🚀

What's New?

🧠 LLMs

• LLM generation now runs on modular TypeScript runners (text, multimodal, Gemma) with a fully featured sampler based on the field-tested llama.cpp one, with support for constrained decoding and fine-grained control over generation. (#1508).
• Re-exported LLMs with context windows up to 16k tokens (#1508).
• enable_thinking chat template option and new MiniCPM5-2B (#1571).
  👉 LLM Chat & Generation

👁️ Computer vision

• Depth estimation: Depth Anything V2 / V3 and ZipDepth (#1552). 👉 Docs
• Image inpainting: MI-GAN (#1573). 👉 Docs
• MediaPipe Face Mesh keypoints (#1469).
• Core ML segmentation models may return an index map (#1463).

🔎 Retrieval

• createVectorStore: in-memory vector store with top-k cosine/dot queries and metadata filters (#1520).
• ColBERT late-interaction embeddings with LFM2.5-ColBERT-350M and a maxSim helper (#1521). 👉 Docs

🗣️ Speech

• Kokoro: Chinese (#1557) and word timestamps (#1495).

⚡ Faster, smaller models

• Vulkan is the Android default for style transfer, YOLO26 (detection / seg / pose), RF-DETR, FastSAM, BlazeFace and SDXS (#1479).
• XNNPACK int8 for Whisper, YOLO26, RF-DETR, FastSAM, MiniLM, mpnet, CLIP, Supertonic and SDXS (#1522, #1530, #1537, #1539, #1542); Core ML text embedders served as int8 (#1465).

📊 Benchmarks

Per-device latency and memory for every registry variant. 👉 Benchmarks

⚠️ Breaking changes

• The @experimental LLM runner API changed (#1508). The chat and generation APIs are unchanged.
• The 15 unquantized (bf16/fp16) XNNPACK LLM variants were removed (#1519).
• 66 unquantized XNNPACK variants with a faster quantized twin are deprecated and will be removed in v1.0.0 (#1519).

🐛 Fixes

All fixes shipped in v0.10.1 – v0.10.7 are included.

Full Changelog: v0.10.0...v0.11.0