vinnj/Qwen3.6-35B-A3B-TypeScript-Edition-GGUF

Qwen3.6-35B TypeScript Edition

Qwen3.6-35B-A3B cut down to the experts it uses for TypeScript, in one 8.8 GB GGUF that fits in memory on a 16 GB Mac and runs directly in stock llama.cpp. Nothing to install from us, nothing to squeeze.

It is a TypeScript-only model: medium to good at backend TypeScript, which is what we tested it on (Node.js and Express routes and middleware, types and generics, async error handling, Jest suites, JavaScript to strict TypeScript). Use it for that and nothing else.

On a Mac mini M4 with 16 GB it writes about 36 tokens a second, gives the same answer every run at temperature 0, and scores 31/40 on our five TypeScript tasks, where the full 22 GB model scores 30/40 in the same mode.

Run it

brew install llama.cpp
curl -L -o ~/Qwen3.6-35B-A3B-TS96-D4.gguf https://huggingface.co/vinnj/Qwen3.6-35B-A3B-TypeScript-Edition-GGUF/resolve/main/Qwen3.6-35B-A3B-TS96-D4.gguf
llama-server -m ~/Qwen3.6-35B-A3B-TS96-D4.gguf -c 16384 --chat-template-kwargs '{"enable_thinking":false}'

Point any OpenAI-compatible tool at http://127.0.0.1:8080/v1. Needs Apple silicon, 16 GB or more, macOS 15 or later. Tested with llama.cpp b11715 and Homebrew llama.cpp 0.7.0.

Check the file: shasum -a 256 must print eafe1005cfda7745aac9f0ab4d9c27d7ee6f9811ee0adce62560f1515d1d954d.

Licence

Apache 2.0. Base model Qwen/Qwen3.6-35B-A3B, cut from unsloth/Qwen3.6-35B-A3B-GGUF. Made by Kraits.