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 quantization of dervig/m51Lab-MiniMax-M2.7-REAP-139B-A10B — the first publicly available REAP-40 % pruned variant of MiniMax-M2.7 — targeting NVIDIA Blackwell (B100 / B200) for native FP4 tensor-core acceleration.
Aspect
Value
Base model
dervig/m51Lab-MiniMax-M2.7-REAP-139B-A10B (BF16)
Quantization
NVFP4A16 (4-bit microscaled floating point weights, FP16 activations)
NVFP4 is NVIDIA's 4-bit floating-point microscaling format introduced with the Blackwell architecture. It uses small block-wise scale factors to maintain quality at extreme compression, and benefits from dedicated FP4 tensor cores on B100/B200 hardware.
Compared to INT4 / AWQ quantization, NVFP4 typically preserves quality better at the same weight budget, particularly on reasoning-heavy workloads. Our REAP-pruned base model is an ideal candidate — the structural pruning has already reduced parameter count, and NVFP4 then packs each remaining weight into 4 bits.
Hardware & deployment
Native FP4 tensor-core acceleration requires Blackwell (B100 / B200). The quantized weights also load and run on Hopper (H100 / H200) and Ampere (A100) via FP4-to-higher-precision upcasting — functional but not at Blackwell speed.
Supported via the compressed-tensors loader in TensorRT-LLM 0.14+ with NVFP4 scheme. Consult NVIDIA's deployment guide for Blackwell-specific kernels.
Quality
Inference quality was validated on the BF16 parent via a 5 / 5 pre-publish smoke test and full HumanEval evaluation (see parent safetensors card). NVFP4A16 is expected to track FP8 / BF16 quality very closely thanks to microscaling — activations remain in FP16 so only weights are compressed.
Systematic NVFP4-on-REAP evaluation is pending; we will update this card if there is community demand.
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, calibration on 3 × 2,048 samples (code / math / tool)
See the parent safetensors card for full architecture, pruning details, evaluation numbers, and the known minor layer-0 bias imperfection.
