3.20.0 (2026-08-11)

Features

• Muse Glimmer support (#639) (adb92f2)
• improve thought segments syntax extraction (#636) (3f686d7)
• expose download speed and ETA for a model downloader (#636) (3f686d7)
• expose the model's architecture directly on the model instance (#636) (3f686d7)
• inspect gpu command: print model parameters count (#639) (adb92f2)

Bug Fixes

• adapt to breaking llama.cpp changes (#636) (3f686d7)
• LlamaContextSequence: make .dispose() return a promise (#636) (3f686d7)
• Qwen chat wrapper auto thought segment opening (#636) (3f686d7)
• optimize checkpoints with auto opening thought segments (#636) (3f686d7)
• simulator model dispose while in use race conditions (#636) (3f686d7)
• Vulkan device memory readings (#636) (3f686d7)
• topP config when using temperature (#639) (adb92f2)
• support more quant labels (#639) (adb92f2)

Shipped with llama.cpp release b10361

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