REAP-pruned models
Configurations whose base checkpoint was pruned with REAP (Router-weighted Expert
Activation Pruning) — the REAP tag. ← all tag kinds
REAP (Cerebras, arXiv:2510.13999, code: CerebrasResearch/reap) is a one-shot, forward-pass-only compression for MoE models: it scores each routed expert by router-gate weight × output magnitude over calibration tokens and drops (does not merge) the lowest-saliency experts in every layer, renormalising the surviving router gates. The result is the same HF architecture with fewer experts — a drop-in for vanilla vLLM and a clean BF16 source for downstream FP8/NVFP4 quantization.
Two flavours appear on this site:
- Self-REAP — models we REAP-pruned (none yet; that page documents how to do it).
- Self-Quantized — models we quantized ourselves, which here means taking Cerebras’s already-REAP’d checkpoints to NVFP4.
REAP (5)
| Configuration | Engine | Quant | Ctx | Conc | Decode tok/s | Status | Completed |
|---|---|---|---|---|---|---|---|
| MiniMax-M2.7-REAP 172B · vLLM · NVFP4 (W4A4) | vLLM | NVFP4 (W4A4) | 65536 | 32 | 111.9 | done | 2026-06-27 19:22 +0800 |
| MiniMax-M2.7-REAP 172B · vLLM · NVFP4 (W4A4) · conc 1 · 160K | vLLM | NVFP4 (W4A4) | 163840 | 1 | 25.4 | done | 2026-06-27 19:22 +0800 |
| MiniMax-M2.5-REAP 139B · vLLM · NVFP4 (W4A16) · conc 1 · 192K | vLLM | NVFP4 (W4A16) | 196608 | 1 | 27.0 | done | 2026-06-27 15:22 +0800 |
| MiniMax-M2.5-REAP 139B · vLLM · NVFP4 (W4A16) | vLLM | NVFP4 (W4A16) | 65536 | 32 | 120.0 | done | 2026-06-26 22:25 +0800 |
| GLM-4.5-Air-REAP 82B · vLLM · NVFP4 (W4A16) | vLLM | NVFP4 (W4A16) | 65536 | 32 | 158.4 | done | 2026-06-26 16:10 +0800 |