✨ Gemma 4 is here! ✨

Read about the release in the blog post

3.19.0 (2026-06-30)

Features

• Gemma 4 support (#591) (5fe6e27) (documentation: Gemma 4)
• riscv64 prebuilt binaries (#615) (e8336a4)
• automatically enable flash attention when optimal
• improve inference performance when a grammar is active
• more precise resource usage estimation
• resource usage capping (documentation: Resource Capping)
• automatically enable or disable mmap depending on the environment
• support Q1_0 quant
• improve stability on unified memory systems
• disable residency sets on macOS by default for better OS responsiveness
• default progressLogs to "stderr" to avoid polluting stdout with logs
• optimized prebuilt binaries for arm architectures

Bug Fixes

• MXFP4_MOE quant name
• Vulkan backend successful load detection even when no devices are available
• CLI: avoid redownloading existing models that consist of multiple parts from a URI
• optimize checkpoints management when using grammar
• improve stability when loading huge models
• reranking result range for Qwen 3 reranker
• adapt to breaking llama.cpp changes

Shipped with llama.cpp release b9842

To use the latest llama.cpp release available, run npx -n node-llama-cpp source download --release latest. (learn more)