Tool routing on-device with a 0.8-billion-parameter model
MLX-LM LoRA on a Qwen3.5-0.8B, GBNF grammars that make schema violations impossible, and smoke-test numbers labeled as smoke-test numbers.
MLX-LM LoRA on a Qwen3.5-0.8B, GBNF grammars that make schema violations impossible, and smoke-test numbers labeled as smoke-test numbers.