How to deploy MiniMax M2.5 on Lambda

TL;DR: token throughput (SGLang)

Benchmark configuration:

vllm bench serve \
  --backend openai-chat \
  --model MiniMaxAI/MiniMax-M2.5 \
  --served-model-name minimax-m25 \
  --dataset-name random \
  --random-input-len 8192 \
  --random-output-len 1024 \
  --num-prompts 512 \
  --max-concurrency 32 \
  --endpoint /v1/chat/completions

See Benchmarking MiniMax-M2.5 for the full command.

Background

MiniMax-M2.5 is a 229B parameter Mixture-of-Experts (MoE) language model with only 10B parameters active per token. Positioned as "the world's first production-level model designed natively for agent scenarios," M2.5 uses 256 experts with 8 activated per token, enabling efficient inference while maintaining frontier-level capabilities.

MiniMax trained using MiniMax's novel Forge framework, an agent-native RL (reinforcement learning) training system that achieves 40x training speedup through prefix-tree merging and windowed FIFO scheduling. M2.5 introduces task completion time optimization, a reward shaping approach that balances intelligence with speed, resulting in 37% faster task completion than its predecessor M2.1.

On benchmarks, M2.5 achieves 80.2% on SWE-Bench Verified (up from M2.1's 74.0%), matching Claude Opus 4.6's speed while costing 10-20x less. The model also demonstrates strong cross-scaffold generalization, performing consistently across different agent frameworks, including Claude Code (80.2%), Droid (79.7%), and OpenCode (76.1%).

Model specifications

Overview

• Name: MiniMax-M2.5
• Author: MiniMaxAI
• Architecture: MoE
• License: MiniMax-Open (Modified MIT)

Specifications

• Total parameters: 229B (10B active per token)
• Context window: 192k tokens
• Languages: English, Spanish, French, German, Italian, Portuguese, Russian, Arabic, Hindi, Korean, Chinese, Japanese, Indonesian, Vietnamese, and Bengali

Recommended Lambda VRAM configuration

• Minimal deployment:
  • 2× B200 (--tp-size 2)
  • 4× H100 (--tp-size 4)

Deployment and benchmarking

Deploying MiniMax-M2.5

MiniMax-M2.5 requires 2× B200 or 4× H100 GPUs to load the model.

1. Launch an instance with 2× B200 or 4× H100 from the Lambda Cloud Console using the GPU Base 24.04 image.
2. Connect to your instance via SSH or JupyterLab terminal. See Connecting to an instance for detailed instructions.
3. Start the SGLang server:

# Use --tp-size 2 for 2× B200, --tp-size 4 for 4× H100
docker run \
    --gpus all \
    -p 8000:8000 \
    --ipc=host \
    -e HF_HOME=/root/.cache/huggingface \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    lmsysorg/sglang:latest \
    python -m sglang.launch_server \
    --host 0.0.0.0 \
    --port 8000 \
    --model-path MiniMaxAI/MiniMax-M2.5 \
    --served-model-name minimax-m25 \
    --tp-size 2 \ # Or 4 for H100
    --tool-call-parser minimax-m2 \
    --reasoning-parser minimax-append-think \
    --trust-remote-code \
    --mem-fraction-static 0.85

This launches an SGLang server with an OpenAI-compatible API on port 8000.

4. Verify the server is running:

curl -X GET http://localhost:8000/v1/models \
  -H "Content-Type: application/json"

You should see MiniMax-M2.5 listed in the response.

Benchmarking MiniMax-M2.5

You can benchmark MiniMax-M2.5 using vllm bench serve. The benchmark results in this article use an 8192 input/1024 output token workload.

Here's a minimal example to run against your server:

vllm bench serve \
  --backend openai-chat \
  --model MiniMaxAI/MiniMax-M2.5 \
  --served-model-name minimax-m25 \
  --dataset-name random \
  --random-input-len 8192 \
  --random-output-len 1024 \
  --num-prompts 512 \
  --max-concurrency 32 \
  --endpoint /v1/chat/completions

Token throughput (2× NVIDIA B200):

Latency details (2× NVIDIA B200):

Next steps

To get started with MiniMax-M2.5, follow the directions above to deploy on Lambda's infrastructure accelerated by NVIDIA. View additional resources about the model below:

• Download the MiniMax-M2.5 weights on Hugging Face.