MiniMax M2.5

Model Overview

Architecture

MiniMax M2.5 uses Lightning Attention, a hybrid approach that splits sequence processing into two components:

• Intra-chunk (linear attention): Processes tokens within a chunk in O(n) time, no KV cache needed
• Inter-chunk (SoftMax attention): Standard attention across chunk boundaries, maintains KV cache

Combined with MoE routing (only 10B of 229B parameters active per token), this creates an efficient model for long-context workloads. The 256-expert pool with 8 active per token provides high capacity with low per-token compute cost.

Implications for NVIDIA HGX B200 deployment:

1. Moderate TP requirement: 10B active params fits well with TP=4 (4 GPUs)
2. Hybrid KV cache: Linear attention layers have no KV cache; SoftMax layers do
3. Good throughput scaling: MoE activates only 4.4% of parameters per token

Quick Start

$ vllm serve MiniMaxAI/MiniMax-M2.5 \
  --tensor-parallel-size 4 \
  --max-model-len 32768 \
  --gpu-memory-utilization 0.90 \
  --trust-remote-code

Or with Docker:

$ docker run --rm --gpus all --ipc=host \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  -p 8000:8000 \
  vllm/vllm-openai:v0.16.0 \
  --model MiniMaxAI/MiniMax-M2.5 \
  --tensor-parallel-size 4 \
  --max-model-len 32768 \
  --gpu-memory-utilization 0.90 \
  --trust-remote-code

Configuration

Memory Usage (NVIDIA HGX B200 Verified)

With TP=4 on FP8:

Performance (NVIDIA HGX B200 Verified)

Benchmark parameters: 2048 input tokens, 512 output tokens, random dataset. TP=4 on 4x NVIDIA HGX B200.

Concurrency Scaling

Zero failed requests across all concurrency levels.

Peak Performance

Scaling Behavior

Throughput scales linearly from concurrency 1 to 512 (~101x increase), then plateaus. The model saturates at approximately 8,800 tok/s sustained output throughput on 4 GPUs. TTFT remains under 100ms up to concurrency 32, making this suitable for interactive applications at moderate load. TPOT stays under 20ms up to concurrency 32.

Test Endpoints

Chat Completion

$ curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "MiniMaxAI/MiniMax-M2.5",
    "messages": [{"role": "user", "content": "Explain Lightning Attention"}],
    "max_tokens": 256
  }'

Text Completion

$ curl http://localhost:8000/v1/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "MiniMaxAI/MiniMax-M2.5",
    "prompt": "The key difference between linear and softmax attention is",
    "max_tokens": 128
  }'

NVFP4 Variant

A community-quantized NVFP4 variant is available for NVIDIA HGX B200:

$ vllm serve lukealonso/MiniMax-M2.5-NVFP4 \
  --tensor-parallel-size 2 \
  --max-model-len 32768 \
  --gpu-memory-utilization 0.90 \
  --trust-remote-code

NVFP4 could potentially halve the GPU count from TP=4 to TP=2. See FP8/NVFP4 Quantization for details.

Known Issues

• vLLM 0.16.0 incompatible: MiniMax M2.5 crashes on vLLM 0.16.0 with RuntimeError: n_group should not be zero for DeepSeekV3 routing. The fused MoE kernel assumes DeepSeek V3-style grouped routing. Use vLLM 0.12.0 for this model. All benchmarks in this guide were run on vLLM 0.12.0. See Troubleshooting for details.
• MoE config warning: vLLM may log warnings about using default MoE config for NVIDIA HGX B200. Performance is still strong — tuned kernel configs for Blackwell are expected in future vLLM releases.
• Custom code: Always pass --trust-remote-code. The Lightning Attention implementation requires custom model code.