Qwen3.5-24B REAP squeezes agentic coding into 16GB

This is a practical community optimization drop, not a new base model, but it materially lowers the barrier to running strong MoE coding models locally.

• –The model trims experts from 256 to 175 while keeping ~3B active parameters per token, targeting better memory efficiency.
• –The recommended IQ4_K_S GGUF is positioned for 16GB-class GPUs, which is the core value proposition here.
• –The author shares full replication assets (REAP fork + Modal scripts), making this useful for other quantizers and pruning experiments.
• –Calibration limits (1024 context and memory pressure during profiling) suggest further quality/perf gains are still possible.