> For the complete documentation index, see [llms.txt](https://docs.clore.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.clore.ai/guides/language-models/minimax-m3.md).

# MiniMax M3 (428B Multimodal)

Deploy MiniMax M3, a 428B natively multimodal MoE with a 1M-token context, on multi-GPU Clore.ai rigs

{% hint style="info" %}
**Status (August 2026):** MiniMax released **M3** on **1 June 2026**, with weights on Hugging Face by 7 June: [MiniMaxAI/MiniMax-M3](https://huggingface.co/MiniMaxAI/MiniMax-M3). **428B total / \~23B active**, **1M-token context**, native **text + image + video** input, under a custom **`minimax-community`** licence (not Apache or MIT — read it before shipping commercially).
{% endhint %}

M3's interesting property is not its size, it is its attention. **MiniMax Sparse Attention (MSA)** cuts per-token compute at 1M context to roughly **1/20** of grouped-query attention, giving MiniMax a claimed **9× faster prefill and 15× faster decode** than M2 at full context. For long-document and long-video work, that is the difference between a model you can afford to fill the window on and one you cannot.

With only \~23B parameters active per token, decode speed on a quantized build is closer to a mid-size model than to a 428B one.

### Key Specs

| Property        | Value                                                            |
| --------------- | ---------------------------------------------------------------- |
| Parameters      | 428B total, \~23B active (MoE)                                   |
| Architecture    | 60 layers, GQA + MiniMax Sparse Attention, native vision encoder |
| Modality        | Text, image, video in → text out                                 |
| Context         | 1,000,000 tokens                                                 |
| License         | `minimax-community` (custom, commercially restricted)            |
| Release Date    | 1 June 2026                                                      |
| Weights         | BF16 854GB · Q2\_K\_XL GGUF 143GB · Q3\_K\_M GGUF 195GB          |
| Primary Tooling | vLLM, SGLang, llama.cpp, Transformers                            |

***

## Requirements

| Build          | Size    | Clore.ai rig                                                       |
| -------------- | ------- | ------------------------------------------------------------------ |
| `UD-IQ1_M`     | 128GB   | 6× RTX 5090 (186GB) — comfortable                                  |
| `UD-Q2_K_XL`   | 143GB   | 6× RTX 5090 or 8× RTX 4090 (192GB)                                 |
| `UD-Q3_K_M`    | 195GB   | 8× RTX 5090 (248GB)                                                |
| `UD-Q4_K_S`    | 248GB   | 10× RTX 5090 (310GB) or 4× RTX PRO 6000 (380GB)                    |
| MXFP8 official | \~430GB | Not on the marketplace → [bare metal](https://clore.ai/bare-metal) |
| BF16 official  | 854GB   | [bare metal](https://clore.ai/bare-metal)                          |

At the last snapshot there were **72 servers with ≥167GB of VRAM (26 free)** and **54 at ≥192GB (14 free)** — enough to place a Q2 or Q3 build, but not so many that you should assume one will be waiting.

{% hint style="warning" %}
Multimodal input needs the vision tower loaded alongside the language model. Budget a few extra GB over the quant size if you plan to feed it images or video, and prefer rigs with plenty of system RAM for video decoding.
{% endhint %}

***

## Deploy with llama.cpp

```bash
huggingface-cli download unsloth/MiniMax-M3-GGUF \
  --include "*UD-Q2_K_XL*" --local-dir /workspace/m3

llama-server -m /workspace/m3/*UD-Q2_K_XL*-00001-of-*.gguf \
  --host 0.0.0.0 --port 8080 \
  -ngl 999 --split-mode layer \
  -c 262144 --flash-attn --no-mmap
```

## Deploy with vLLM

```bash
vllm serve MiniMaxAI/MiniMax-M3 \
  --tensor-parallel-size 8 \
  --trust-remote-code \
  --max-model-len 262144 \
  --gpu-memory-utilization 0.92 \
  --enable-expert-parallel
```

Pre-quantized checkpoints: [`MiniMaxAI/MiniMax-M3-MXFP8`](https://huggingface.co/MiniMaxAI/MiniMax-M3-MXFP8) (official), [`nvidia/MiniMax-M3-NVFP4`](https://huggingface.co/nvidia/MiniMax-M3-NVFP4) for Blackwell, [`cyankiwi/MiniMax-M3-AWQ-INT4`](https://huggingface.co/cyankiwi/MiniMax-M3-AWQ-INT4) for AWQ stacks.

***

## Clore.ai GPU Recommendations

| Setup           | Total VRAM | Build        | Clore.ai cost       |
| --------------- | ---------- | ------------ | ------------------- |
| 6× RTX 5090     | 186GB      | Q2\_K\_XL    | \~$1.50–2.60/hr     |
| 8× RTX 4090     | 192GB      | Q2\_K\_XL    | \~$1.12–3.36/hr     |
| **8× RTX 5090** | **248GB**  | **Q3\_K\_M** | **\~$2.00–3.50/hr** |
| 4× RTX PRO 6000 | 380GB      | Q4\_K\_S     | \~$5.00/hr          |

***

## Use Cases

* **Long-video understanding** — native video input plus the sparse-attention cost curve
* **Million-token document analysis** — contracts, codebases, archives in a single pass
* **Agentic coding and "cowork" tasks** — what MiniMax optimised M3 for
* **Cheap decode at frontier scale** — 23B active parameters means tokens come out fast for the model's size
* **Multimodal RAG** — one model for text, screenshots and video frames

{% hint style="warning" %}
**Check the licence before commercial use.** M3 ships under `minimax-community`, not an OSI licence. If you need unrestricted terms, [GLM-5.2](/guides/language-models/glm-5-2.md) is MIT and [Qwen3.8-27B](/guides/language-models/qwen38-27b.md) is Apache 2.0.
{% endhint %}

***

## Troubleshooting

| Issue                            | Fix                                                                                  |
| -------------------------------- | ------------------------------------------------------------------------------------ |
| `trust_remote_code` errors       | M3 ships custom modelling code — pass `--trust-remote-code` in vLLM                  |
| Sparse attention kernels missing | Update vLLM/SGLang; MSA support landed after the June release                        |
| Video input fails                | Check the rig has enough system RAM and CPU for decoding; video is decoded host-side |
| OOM at 1M context                | The KV cache is smaller than GQA but not free — start at 262K and work up            |
| Slow prefill                     | Enable chunked prefill; MSA's advantage shows at long context, not on short prompts  |

***

## Next Steps

* **Predecessor:** [MiniMax M2.7](/guides/language-models/minimax-m27.md) — the April coding MoE
* **MIT alternative at similar scale:** [GLM-5.2](/guides/language-models/glm-5-2.md)
* **Single-card alternative:** [Qwen3.8-27B](/guides/language-models/qwen38-27b.md)
* **Rig sizing:** [GPU Pricing & Availability](/guides/getting-started/pricing.md)

### Links

* [MiniMax M3 on Hugging Face](https://huggingface.co/MiniMaxAI/MiniMax-M3) · [GGUF](https://huggingface.co/unsloth/MiniMax-M3-GGUF)
* [Technical report (arXiv 2606.13392)](https://arxiv.org/abs/2606.13392) · [MSA repo](https://github.com/MiniMax-AI/MSA)
* [MiniMax-M3 GitHub](https://github.com/MiniMax-AI/MiniMax-M3)
* **Rent a GPU:** [RTX 5090](https://clore.ai/rent-5090.html) · [RTX 4090](https://clore.ai/rent-4090.html) · [Marketplace](https://clore.ai/marketplace)


---

# Agent Instructions
This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com.

## Querying This Documentation
If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question.

Perform an HTTP GET request on the current page URL with the `ask` query parameter, and the optional `goal` query parameter:

```
GET https://docs.clore.ai/guides/language-models/minimax-m3.md?ask=<question>&goal=<endgoal>
```

`ask` is the immediate question: it should be specific, self-contained, and written in natural language.
`goal` is optional and describes the broader end goal you are ultimately trying to accomplish on behalf of the user. GitBook uses it to tailor the answer towards what is most useful for that goal.

The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.

Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.