NOTE: MiniMax M2.5 is still my daily driver. While Qwen3.5-122B is no slouch, Minimax still delivers consistently, and with a token throughput that allows me to get a lot more done. For a model with 228B parameters, I never expected it to be this good on my hardware, particularly when quantized at Q2. You never know until you try, right? A lot of people have setups that work for them, each person is different, but I am always going back to Minimax. I've tried it in Hermes Agent, but honestly Minimax + Opencode is my winning combo. I've connected it to our "open brain" knowledge base, and so I have memory and everything baked in. Other models and Hermes don't reliably access it, but Minimax+Opencode get it right. Other models we've tried recently were Mistral Small-4-119B and MiMo-V2-Flash—just couldn't get their configs right, and once they were running, things were rough and slow. Call it user error but as of today, Minimax still stands alone, in my view. This post documents MiniMax M2.5 running on the Ironhorse rig with six RTX 3090s, including the llama.cpp upgrade that almost broke everything and the fix that made it work properly again.

This blog post was generated using MiniMax M2.5 running through Opencode.

The Backstory

Hardware and Model

System Specs

• GPUs: 6x NVIDIA RTX 3090 (24GB VRAM each, 144GB total)
• CPU: AMD Threadripper 5995WX
• RAM: 64GB DDR4-3200 ECC
• PCIe: Full 4.0 16x speed on all GPUs (no NVLink, no P2P)
• Framework: llama.cpp server (latest master, ~b8400)

gratuitous GPU porn

Model Details

MiniMax M2.5 is a formidable model:

llama.cpp Configuration: The Critical Details

Startup Script (minimax-m2.5.sh)

#!/bin/bash

MODEL_PATH="/home/tomwest/models/minimax-m2.5/MiniMax-M2.5-UD-IQ2_M-00001-of-00003.gguf"
LLAMA_SERVER="$HOME/llama.cpp/build/bin/llama-server"
LOG_FILE="$HOME/models/minimax-m2.5.log"

if pgrep -x "llama-server" > /dev/null; then
    echo "Killing existing llama-server processes..."
    pkill -x "llama-server"
    sleep 1
fi

echo "Starting MiniMax-M2.5 IQ2_M on port 8000..."
nohup "$LLAMA_SERVER" \
    -m "$MODEL_PATH" \
    --host 0.0.0.0 \
    --port 8000 \
    -ngl 999 \
    --flash-attn on \
    -c 204800 \
    --parallel 1 \
    --split-mode layer \
    -ts 0.8,1,1,1,1,1 \
    --temp 1.0 \
    --top-p 0.95 \
    --top-k 40 \
    --jinja \
    "$@" > "$LOG_FILE" 2>&1 &

Key Parameters

• -ngl 999: Offload all 80 layers to GPU
• -c 204800: Full 200K context window
• --parallel 1: Single slot—all KV cache budget to one user
• --flash-attn on: Enables Flash Attention—critical for hybrid GDN models
• --split-mode layer: Layer-wise GPU split (vs row-wise)
• -ts 0.8,1,1,1,1,1: Tensor split ratio—GPU0 derated for COSMIC desktop
• --temp 1.0 --top-p 0.95 --top-k 40: Recommended sampling settings
• --jinja: Native Jinja template engine

Benchmark Results

We ran comprehensive benchmarks across context lengths (4K to 198K) at 350W power limit per GPU. Here are the results from running through Opencode:

Performance Visualization

The following charts visualize the performance characteristics across different context lengths:

Prefill Speed vs Context Length

Prefill phase performance across context lengths (4K to 198K). Four panes: throughput vs context, efficiency vs context, energy efficiency, and power range.

Decode Speed vs Context Length

Decode phase performance across context lengths (4K to 198K). Six panes: throughput vs context, efficiency vs context, energy efficiency, power range, TTFT, and efficiency vs power.

Memory Analysis

The IQ2_M quantized model is ~107GB, fitting in 144GB VRAM with room for KV cache:

opencode Integration

Added to ~/.config/opencode/opencode.json under the llama-cpp provider:

{
  "MiniMax-M2.5-UD-IQ2_M": {
    "name": "MiniMax M2.5",
    "contextLength": 204800
  }
}

MiniMax M2.5 vs M2.1: A Comparison

Would I Recommend This?

TL;DR: Yes. MiniMax M2.5 in IQ2_M is a strong choice for local inference. The 200K context window is fully usable, the reasoning capabilities are solid, and with the right llama.cpp configuration, it delivers 60+ tokens/sec decode speed. While Qwen3.5-122B is no slouch, Minimax still delivers consistently and with a token throughput that allows me to get a lot more done. For a model with 228B parameters, I never expected it to be this good on my hardware, particularly when quantized at Q2. You never know until you try, right?

A lot of people have setups that work for them, each person is different. Other models we've tried recently—Mistral Small-4-119B and MiMo-V2-Flash—couldn't get their configs right, and once they were running, things were rough and slow. Call it user error, but as of today, Minimax still stands alone.

MiniMax M2.5 IQ2_M on six RTX 3090s demonstrates that local AI doesn't mean compromise. With ~10B active parameters per token and 65 tokens/sec decode speed, you get interactive performance with strong reasoning. The configuration is stable when done right—and now you know how to do it right.