What happened: The initial upload (2026-04-15) used ignore=["lm_head"] in the llm-compressor recipe, which meant the 62 MoE routers (block_sparse_moe.gate) got quantized along with the expert weights. vLLM's MiniMax-M2 loader expects an unquantized ReplicatedLinear router and fails at engine-init with:
This is a hard load failure — the engine never initializes, so no tokens are generated. (The earlier "degraded output" framing understated the severity.)
Root cause: Missing MoE-aware entries in the llm-compressor ignore list. The correct pattern (per saricles/MiniMax-M2.5-REAP-139B-A10B-NVFP4-GB10):
Fix: This variant was re-rolled 2026-04-16 with the corrected recipe. quantization_config.ignore now lists all 62 per-layer router gates alongside lm_head.
Verification:config.json on this repo now contains 62 model.layers.N.block_sparse_moe.gate entries in the ignore list. Loaders should open the model without the KeyError above.
Credit: Thanks to the community user who reported this first on the NVFP4-GB10 DGX Spark load. The saricles reference repo was invaluable for confirming the exact pattern.
Unaffected variants (no re-roll needed): BF16 safetensors, all GGUF quantizations.
NVFP4 W4A4 (FP4 weights and activations) of dervig/m51Lab-MiniMax-M2.7-REAP-139B-A10B — the first publicly available REAP-40 % pruned variant of MiniMax-M2.7 — specifically targeting GB10 (NVIDIA DGX Spark / Project Digits, SM12.1) and Blackwell FP4-native workloads.
Aspect
Value
Base model
dervig/m51Lab-MiniMax-M2.7-REAP-139B-A10B (BF16)
Quantization
NVFP4 (microscaled FP4 for both weights and activations — W4A4)
Dequanting W4A16 to BF16 at load costs ~260 GB — exceeds 128 GB unified memory
This W4A4 variant is the canonical format for GB10 and routes through the native FP4 kernel path with Marlin fallback. Follows the established saricles/MiniMax-M2.5-REAP-172B-A10B-NVFP4-GB10 convention.
Inference quality validated on the BF16 parent via a 5 / 5 pre-publish smoke test and full HumanEval evaluation (see parent safetensors card). W4A4 quantization has more aggressive compression than W4A16 — activation quantization adds a modest quality delta vs FP8 or the W4A16 NVFP4 — typically 1-3 % on reasoning benchmarks for this class of model. For maximum quality on Blackwell, prefer the FP8 or W4A16 NVFP4 variants; for GB10 deployment where 128 GB memory is the binding constraint, this W4A4 variant is the canonical choice.
Base model summary
Property
Value
Architecture
MoE, 62 layers, 154 experts (pruned from 256), top-8 routing
Active parameters / token
~10 B
Total parameters
~139 B
Max position embeddings
196,608
Vocabulary size
200,064
Pruning
REAP 40 %, seed 42
See the parent safetensors card for full architecture, pruning details, and known minor layer-0 bias imperfection.
