The 80B Qwen3-Coder-Next on a 16 GB Mac, about 13 tok/s, fully offline. A mixture-of-experts inference engine for Apple silicon.
Rust
#!/bin/bash
# Every model file given loads and answers a one-word question; a file of an unsupported
# architecture is refused with a clear message, never a crash.
# tests-mac/load_models.sh TOKENIZER MODEL.gguf [MODEL.gguf ...] [--unsupported FILE.gguf]
set -u
BIN=${COACHWHIP_BIN:-./target/release/coachwhip}
TOKENIZER=${1:?tokenizer.json}; shift
fail=0
while [ $# -gt 0 ]; do
if [ "$1" = "--unsupported" ]; then
out=$("$BIN" --model "$2" --tokenizer "$TOKENIZER" --max-tokens 4 --prompt "Hi" 2>&1)
code=$?
if [ $code -ne 0 ] && echo "$out" | grep -qi "support"; then echo "refused $(basename "$2")"
else echo "FAIL $(basename "$2"): exit $code, no clear refusal"; fail=1; fi
shift 2; continue
fi
out=$("$BIN" --model "$1" --tokenizer "$TOKENIZER" --temperature 0 --max-tokens 8 --prompt "Say OK." 2>&1)
code=$?
if [ $code -eq 0 ] && echo "$out" | grep -q "model ready"; then echo "loads $(basename "$1")"
else echo "FAIL $(basename "$1"): exit $code"; echo "$out" | tail -3; fail=1; fi
shift
done
[ $fail = 0 ] && echo "load_models: PASS" || echo "load_models: FAIL"
exit $fail