v3.18.0

[github-actions[bot]]

3.18.0 (2026-03-15)

Features

• automatic checkpoints for models that need it (#573) (c641959)
• QwenChatWrapper: Qwen 3.5 support (#573) (c641959)
• inspect gpu command: detect and report missing prebuilt binary modules and custom npm registry (#573) (c641959)

Bug Fixes

• resolveModelFile: deduplicate concurrent downloads (#570) (cc105b9)
• correct Vulkan URL casing in documentation links (#568) (5a44506)
• Qwen 3.5 memory estimation (#573) (c641959)
• grammar use with HarmonyChatWrapper (#573) (c641959)
• add mistral think segment detection (#573) (c641959)
• compress excessively long segments from the current response on context shift instead of throwing an error (#573) (c641959)
• default thinking budget to 75% of the context size to prevent low-quality responses (#573) (c641959)

---

Shipped with llama.cpp release b8352

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

---

This discussion was created from the release v3.18.0.