---
base_model:
  - cerebras/MiniMax-M2.5-REAP-172B-A10B
license: mit
license_name: modified-mit
library_name: transformers
pipeline_tag: text-generation
language:
  - en
  - zh
tags:
  - minimax
  - nvfp4
  - 4-bit
  - quantized
  - compressed-tensors
  - vllm
  - DGX-Spark
  - GB10
  - MoE
  - REAP
---

# MiniMax-M2.5-REAP-172B-A10B-NVFP4-GB10

NVFP4 quantization of [cerebras/MiniMax-M2.5-REAP-172B-A10B](https://huggingface.co/cerebras/MiniMax-M2.5-REAP-172B-A10B) for NVIDIA DGX Spark (GB10).

The base model is a [Cerebras REAP](https://www.cerebras.ai/blog/reap) (Router-weighted Expert Activation Pruning) variant of [MiniMaxAI/MiniMax-M2.5](https://huggingface.co/MiniMaxAI/MiniMax-M2.5). REAP uniformly prunes experts from 256 → 192, reducing total parameters from 230B to 172B while maintaining near-identical performance.

## Model Details

| | |
|---|---|
| **Base Model** | [cerebras/MiniMax-M2.5-REAP-172B-A10B](https://huggingface.co/cerebras/MiniMax-M2.5-REAP-172B-A10B) |
| **Original Model** | [MiniMaxAI/MiniMax-M2.5](https://huggingface.co/MiniMaxAI/MiniMax-M2.5) (230B) |
| **Architecture** | MiniMaxM2ForCausalLM (MoE, 192 experts, 8 active per token) |
| **Total Parameters** | 172B |
| **Active Parameters** | 10B per token |
| **Quantization** | NVFP4 (4-bit floating point), all layers including self_attn |
| **Format** | compressed-tensors (safetensors), 20 shards |
| **Size on Disk** | 99 GB |
| **Context Length** | 196,608 tokens (~192K) |
| **License** | Modified MIT (inherited from [Cerebras REAP](https://huggingface.co/cerebras/MiniMax-M2.5-REAP-172B-A10B)) |

## Performance (Single NVIDIA DGX Spark — GB10, 128 GB)

Benchmarked with [llama-benchy](https://github.com/neuralmagic/llama-benchy).

| Metric | Value |
|--------|-------|
| **Decode throughput** | 27–29 tok/s |
| **Prefill (512 tokens)** | 920 tok/s |
| **Prefill (4096 tokens)** | 1,916 tok/s |
| **TTFT (512 tokens)** | 490 ms |
| **Max context (gpu_mem_util=0.93)** | 65,536 tokens |
| **KV cache capacity** | ~127K tokens |

Effective throughput with a large system prompt (~23K tokens): ~21 tok/s.

## Quantization Details

- **Method:** Post-training quantization via [LLM Compressor](https://github.com/vllm-project/llm-compressor)
- **Calibration Dataset:** [HuggingFaceH4/ultrachat_200k](https://huggingface.co/datasets/HuggingFaceH4/ultrachat_200k) (train_sft split)
- **Calibration Samples:** 64
- **Max Sequence Length:** 2048 tokens
- **Ignore List:** `lm_head`, `model.embed_tokens`, `re:.*block_sparse_moe\.gate$`
- **Environment:** `LLMCOMPRESSOR_MOE_CALIBRATE_ALL_EXPERTS=1`
- **Hardware Used:** NVIDIA DGX Spark with CPU offloading + swap (~4.7 hours)

## Running on a Single DGX Spark

This is exactly how we run and benchmark this model. One DGX Spark, nothing else.

**Docker image:** [`avarok/dgx-vllm-nvfp4-kernel:v23`](https://hub.docker.com/r/avarok/dgx-vllm-nvfp4-kernel) (vLLM 0.16.0-rc2, CUDA 13.0, SM 12.1)

**Download the model:**

```bash
huggingface-cli download saricles/MiniMax-M2.5-REAP-172B-A10B-NVFP4-GB10 \
  --local-dir /opt/huggingface/models/MiniMax-M2.5-REAP-172B-NVFP4
```

**Launch:**

```bash
docker run -d --name minimax --gpus all --ipc=host \
  -v /opt/huggingface/models/MiniMax-M2.5-REAP-172B-NVFP4:/models/MiniMax-M2.5-REAP-172B-NVFP4 \
  -p 8000:8000 \
  -e VLLM_NVFP4_GEMM_BACKEND=marlin \
  -e VLLM_TEST_FORCE_FP8_MARLIN=1 \
  -e VLLM_USE_FLASHINFER_MOE_FP4=0 \
  -e VLLM_MARLIN_USE_ATOMIC_ADD=1 \
  -e MODEL=/models/MiniMax-M2.5-REAP-172B-NVFP4 \
  -e PORT=8000 \
  -e MAX_MODEL_LEN=65536 \
  -e GPU_MEMORY_UTIL=0.93 \
  -e "VLLM_EXTRA_ARGS=--trust-remote-code --kv-cache-dtype fp8 --attention-backend flashinfer --enable-auto-tool-choice --tool-call-parser minimax_m2 --reasoning-parser minimax_m2_append_think" \
  avarok/dgx-vllm-nvfp4-kernel:v23
```

Model takes ~3–4 minutes to load. Verify it's ready:

```bash
curl http://localhost:8000/v1/models
```

**Test it:**

```bash
curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "MiniMax-M2.5-REAP-172B-NVFP4",
    "messages": [{"role": "user", "content": "Hello!"}],
    "temperature": 1.0,
    "top_p": 0.95,
    "top_k": 40,
    "min_p": 0.01,
    "max_tokens": 512
  }'
```

### What the env vars do

| Variable | Why |
|----------|-----|
| `VLLM_NVFP4_GEMM_BACKEND=marlin` | Use Marlin kernels for FP4 GEMM (FlashInfer JIT fails on Spark) |
| `VLLM_TEST_FORCE_FP8_MARLIN=1` | Required for Marlin backend activation |
| `VLLM_USE_FLASHINFER_MOE_FP4=0` | Disable FlashInfer for MoE FP4 (crashes with JIT ninja build) |
| `VLLM_MARLIN_USE_ATOMIC_ADD=1` | Atomic adds for Marlin (stability on GB10) |
| `GPU_MEMORY_UTIL=0.93` | 0.95 OOMs on Spark; 0.93 is the safe max |
| `--kv-cache-dtype fp8` | FP8 KV cache saves memory, enables ~127K token capacity |
| `--attention-backend flashinfer` | FlashInfer for attention (not MoE) — works fine |

### Notes

- `gpu_memory_utilization=0.95` will OOM. Use 0.93.
- The model serves on port 8000 with an OpenAI-compatible API.
- Prefix caching is enabled by default in vLLM 0.16+.
- With 65K context, you get roughly 2 concurrent sessions with large system prompts.
- Tool calling requires `--enable-auto-tool-choice --tool-call-parser minimax_m2`.

### Recommended Sampling Parameters

Per [MiniMax documentation](https://huggingface.co/MiniMaxAI/MiniMax-M2.5):

```json
{
  "temperature": 1.0,
  "top_p": 0.95,
  "top_k": 40,
  "min_p": 0.01
}
```

## Acknowledgments

- Base model by [MiniMax](https://huggingface.co/MiniMaxAI)
- REAP sparse-inference pruning by [Cerebras](https://huggingface.co/cerebras) ([paper](https://arxiv.org/abs/2510.13999))
- Quantization tooling by [vLLM / LLM Compressor](https://github.com/vllm-project/llm-compressor)