MiniMax M3 is one of the more ambitious open-weights releases of 2026: 428 billion parameters, a one-million-token context window, and native image and video input trained in from the start rather than bolted on afterwards.
It is also genuinely large. This guide covers what M3 is, what it realistically takes to run, and how to deploy it on Runpod without discovering the hardware requirements the expensive way.
What is MiniMax M3?
M3 is an open-weights Mixture-of-Experts model from the Shanghai lab MiniMax, announced on 1 June 2026 with weights landing on Hugging Face the following week.
- Parameters: 428B total, roughly 23B active per token
- Context: 1 million tokens
- Modality: native image and video input, trained on mixed text, image, and video from the start
- Focus: reasoning, agentic coding, and multimodality in a single model
MiniMax positions it as the first open-weights release combining all three of those capabilities, and it reports 59.0% on SWE-Bench Pro. The multimodal training shows up in unusual places: the model can work from screenshots well enough to operate a desktop, which is a different capability profile from most models at this scale.
Verify the license terms on the Hugging Face repository before building commercially. Open weights and permissive licensing are not the same thing, and it is worth five minutes to check rather than assume.
Why run MiniMax M3 on Runpod?
The hardware requirement is the whole story. At 428B parameters, M3 is firmly multi-GPU. This is not a model anyone evaluates on a workstation.
Multimodal input needs headroom. Image and video tokens consume context quickly. A video clip can occupy a substantial share of your window before any text arrives, so memory planning matters more than with a text-only model.
Evaluation should be small relative to commitment. Finding out whether M3's multimodal capability suits your workload takes a few hours of multi-GPU time on Runpod. Building the equivalent capacity to find out does not.
GPU requirements for MiniMax M3
All 428B parameters must be resident even though only about 23B activate per token. Sparse activation buys speed, not storage.
- 4-bit quantized: roughly 240 GB. A single B300 (288 GB) at $6.94/hr, or 2x B200 at $5.98/hr each on Community Cloud.
- 8-bit quantized: roughly 470 GB. Multi-GPU, 2x B300 or 3x B200.
- bf16 full precision: beyond 850 GB. Multi-node Clusters.
Add meaningful headroom on top of these if you intend to use long context or feed video, because KV cache at 1M tokens is substantial in its own right.
The most economical entry point is a 4-bit build on a single B300. Keeping everything on one GPU avoids tensor parallelism overhead entirely, which matters more than the raw price difference against two smaller cards.
Step 1: Create a Runpod account
Go to runpod.io and sign up with email, GitHub, or Google. Add credits or a payment method, and be aware that B300-class hardware accrues faster than the GPUs in most tutorials.
Step 2: Provision storage first
Create a network volume of at least 500 GB before deploying anything, and mount it at /workspace. At M3's size, downloading weights takes long enough that doing it twice is a genuine waste of money.
Step 3: Deploy
- Open the Pods section and select a B300, or a multi-GPU B200 configuration.
- Attach the network volume you created.
- Choose a vLLM or PyTorch template.
- Set container disk to at least 100 GB separately from the volume.
- Deploy.
The pod comes up in under 30 seconds. Weight download and load will take considerably longer on first run.
Step 4: Serve and test
Set max model length well below 1M for your first deployment. Something like 32k is enough to confirm the model works, and KV cache at full context will consume memory before you have sent a single request.
Test text generation first, then multimodal input separately. If something fails, you want to know which capability broke rather than debugging both at once.
Pro tips
Test text before multimodal. Getting a 428B model serving at all is the harder half. Confirm that, then add images and video.
Budget context for media. Video and images consume tokens at a rate that surprises people the first time. Measure how much of your window a typical input actually uses before designing around a number.
Single GPU beats two where possible. If a 4-bit build fits on one B300, that is usually preferable to splitting across two B200s. Tensor parallelism has overhead, and avoiding it is worth a modest price difference.
Check the license before you build. Open weights do not automatically mean unrestricted commercial use. Read the terms on the model repository.
Stop your pod when you are done. Runpod bills by the second, and multi-GPU configurations left running overnight add up quickly.
Wrapping up
M3 is a serious model with a serious hardware requirement. The combination of frontier-scale reasoning, a million-token window, and native video understanding in open weights is genuinely uncommon, and if any of that maps to a problem you have, it is worth the multi-GPU deployment to find out. If it does not, a smaller model will serve you better and cost far less.
Ready to try it? Deploy a B300 pod with a network volume attached, and start with a quantized build at moderate context.
FAQ
What GPU do I need to run MiniMax M3?
A 4-bit quantized build needs roughly 240 GB, which means a single B300 at $6.94/hr or two B200s. Eight-bit roughly doubles that, and full bf16 precision is multi-node cluster territory.
What does 428B with 23B active mean?
The model holds 428 billion parameters but routes each token through only about 23 billion of them. Generation speed resembles a much smaller model, but memory must accommodate the full parameter set.
Can MiniMax M3 process video?
Yes. M3 was trained on mixed text, image, and video from the start rather than having vision added afterwards, and it accepts both image and video input natively. It can work from screenshots well enough to operate a desktop interface.
Is MiniMax M3 free for commercial use?
The weights are openly available on Hugging Face, but check the specific license terms on the model repository before commercial deployment. Open weights and permissive licensing are not the same thing.
Should I use M3 or a smaller model?
Use M3 if you need the combination of long context, multimodal input, and agentic reasoning in one model. If you only need strong text generation, a smaller model such as Qwen3.6 or GLM-4.7-Flash will cost dramatically less to serve.
