> ## Documentation Index
> Fetch the complete documentation index at: https://platform.minimax.io/docs/llms.txt
> Use this file to discover all available pages before exploring further.
# Self-host MiniMax-M3
> Run the experimental MiniMax-M3 SGLang baseline, verify the service, and plan a secure self-hosted deployment.
MiniMax-M3 is a native multimodal Mixture-of-Experts model with approximately 428B total parameters, approximately 23B activated parameters per token, and a 1M-token model context limit. This page covers **server-class self-hosting with SGLang**, not consumer-device local inference.
**Deployment status: Experimental.** As reviewed on August 26, 2026, MiniMax-M3 support is still delivered through an SGLang development image rather than a tagged SGLang release. Pin both the model revision and image digest below, validate them on your workload, and do not treat this recipe as a production SLA.
## Status and reference baseline
| Item | Reference value |
| :------------------------ | :--------------------------------------------------------------------------------------------------- |
| Deployment status | Experimental |
| Last documentation review | August 26, 2026 |
| Reference hardware | 8 × NVIDIA B200, single node, TP 8 |
| Weights | `MiniMaxAI/MiniMax-M3-MXFP8` |
| Model revision | `c5454eb03678d8710e54a4e0fc681b9f3b4a3dba` |
| Weight repository size | Approximately 444 GB at the pinned revision |
| SGLang image | `lmsysorg/sglang@sha256:de63ac56df5d7b064451e21147eaab89634a02332d830ca8c01cb8c033b3a78f` |
| Runtime | Linux and Docker with the NVIDIA Container Toolkit; Python and CUDA are supplied by the pinned image |
| Validation source | [SGLang MiniMax-M3 Cookbook](https://docs.sglang.io/cookbook/autoregressive/MiniMax/MiniMax-M3) |
The official sources do not publish a minimum host-memory requirement, peak GPU-memory measurement, minimum NVIDIA driver version, or complete production disk margin for this recipe. Confirm driver compatibility with the CUDA runtime in the pinned image and provision more disk than the weight repository size before deployment.
The complete weights must be loaded or sharded. “23B activated parameters” does not mean that the deployment only needs memory for 23B parameters.
## Hardware configurations
The following table reports SGLang's current support matrix. “Validated” means validated by the SGLang Cookbook, not independently benchmarked by MiniMax in this document.
| Hardware | Weights | GPUs / TP | SGLang status | Weight storage | Peak GPU memory | Host memory | Topology |
| :------------------ | :----------------------------------- | :-------- | :------------------------------------------- | :------------------- | :-------------- | :------------ | :---------- |
| NVIDIA B200 | MXFP8 | 8 / 8 | Validated; reference baseline | Approximately 444 GB | Not published | Not published | Single node |
| NVIDIA B300 | MXFP8 | 4 / 4 | Validated | Approximately 444 GB | Not published | Not published | Single node |
| NVIDIA GB300 | MXFP8 | 4 / 4 | Validated | Approximately 444 GB | Not published | Not published | Single node |
| NVIDIA GB200 | MXFP8 | 4 / 4 | Inferred by SGLang; not directly benchmarked | Approximately 444 GB | Not published | Not published | Single node |
| NVIDIA H200 | BF16 | 8 / 8 | Validated | Approximately 854 GB | Not published | Not published | Single node |
| AMD MI355X | MXFP8 | 8 / 8 | Validated for text workloads | Approximately 444 GB | Not published | Not published | Single node |
| AMD MI300X | MXFP8 converted at load to block FP8 | 8 / 8 | Validated for text workloads | Approximately 444 GB | Not published | Not published | Single node |
| AMD MI350X / MI325X | MXFP8 | 8 / 8 | Inferred from the same-architecture GPU | Approximately 444 GB | Not published | Not published | Single node |
Use the [SGLang configuration generator](https://docs.sglang.io/cookbook/autoregressive/MiniMax/MiniMax-M3#hw=b200\&variant=default\&quant=mxfp8\&strategy=balanced\&nodes=single) for the exact B300, GB-series, H200, or AMD command. Do not substitute those commands into this B200 baseline without revalidation.
## Quickstart: 8 × B200
Before starting, verify that the host has enough free disk for the approximately 444 GB checkpoint, the container image, and caches.
### 1. Pull the pinned runtime
```bash theme={null}
docker pull lmsysorg/sglang@sha256:de63ac56df5d7b064451e21147eaab89634a02332d830ca8c01cb8c033b3a78f
```
### 2. Verify GPU access
The output must list all eight B200 GPUs:
```bash theme={null}
docker run --rm --gpus all \
lmsysorg/sglang@sha256:de63ac56df5d7b064451e21147eaab89634a02332d830ca8c01cb8c033b3a78f \
nvidia-smi
```
### 3. Start the server
```bash theme={null}
docker run --rm --name minimax-m3 \
--gpus all \
--shm-size 32g \
--ipc=host \
-p 127.0.0.1:30000:30000 \
-v ~/.cache/huggingface:/root/.cache/huggingface \
lmsysorg/sglang@sha256:de63ac56df5d7b064451e21147eaab89634a02332d830ca8c01cb8c033b3a78f \
sglang serve \
--trust-remote-code \
--model-path MiniMaxAI/MiniMax-M3-MXFP8 \
--revision c5454eb03678d8710e54a4e0fc681b9f3b4a3dba \
--reasoning-parser auto \
--tool-call-parser auto \
--tp 8 \
--attention-backend fa4 \
--mm-attention-backend flashinfer_cudnn \
--moe-runner-backend deep_gemm \
--chunked-prefill-size 8192 \
--mem-fraction-static 0.65 \
--host 0.0.0.0 \
--port 30000
```
The process listens on all interfaces **inside the container**, but `-p 127.0.0.1:30000:30000` publishes it only on the host loopback interface. The first start downloads approximately 444 GB of weights and may take a long time, depending on storage and network throughput.
### 4. Check readiness
Run this command in another terminal after the server log reports that it is ready:
```bash theme={null}
curl --fail --silent --show-error http://127.0.0.1:30000/health
```
A successful check exits with status `0`. If it fails, keep the server process running and inspect its logs before sending inference requests.
### 5. Verify text generation
```bash theme={null}
curl http://127.0.0.1:30000/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "MiniMaxAI/MiniMax-M3-MXFP8",
"messages": [
{"role": "user", "content": "Explain sparse attention in one sentence."}
],
"temperature": 1.0,
"top_p": 0.95,
"max_tokens": 1024
}'
```
A successful response contains a non-empty `choices[0].message.content`. When the model emits a reasoning trace, SGLang returns it separately in `choices[0].message.reasoning_content`.
### 6. Verify image input
```bash theme={null}
curl http://127.0.0.1:30000/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "MiniMaxAI/MiniMax-M3-MXFP8",
"messages": [{
"role": "user",
"content": [
{"type": "text", "text": "Describe this image in one sentence."},
{"type": "image_url", "image_url": {"url": "https://raw.githubusercontent.com/sgl-project/sglang/2511743bd784e69e5a81ca3d926a000711dae4ab/examples/assets/example_image.png"}}
]
}],
"temperature": 1.0,
"top_p": 0.95,
"max_tokens": 1024
}'
```
This example uses an image maintained by the SGLang project. In production, accept only trusted or allowlisted remote URLs, or send a Base64 data URI after validating the file at your gateway.
## Capability boundary
| Capability | Model-native capability | This B200 SGLang recipe | Notes |
| :-------------------- | :---------------------- | :----------------------------------------------------- | :----------------------------------------------------------------------------------------------------- |
| Text input and output | Supported | Validated by SGLang | OpenAI-compatible chat completions |
| Reasoning | Supported | Validated by SGLang | Requires `--reasoning-parser auto` |
| Tool calls | Supported | Validated by SGLang | Requires `--tool-call-parser auto`; single, parallel, object, and array arguments are covered upstream |
| Image input | Supported | URL and Base64 inputs validated by SGLang | Blackwell requires `--mm-attention-backend flashinfer_cudnn` |
| Video input | Supported by the model | Not validated in this SGLang recipe | Do not advertise video serving from this recipe without your own validation |
| Context length | Up to 1M tokens | SGLang reports validation from 1K to 128K input tokens | 1M is a model limit, not the validated production default for this configuration |
On AMD, SGLang has validated text chat, reasoning separation, and tool calling. Vision has not been exercised on the ROCm path.
## Published performance data
SGLang reports the following serving measurements. These are upstream reference results, not MiniMax performance guarantees.
| Hardware | Weights / TP | Workload | Mean TTFT | Mean TPOT | Throughput per GPU |
| :------- | :----------- | :----------------------------------------------------------- | :-------- | :-------- | :----------------- |
| 8 × B200 | MXFP8 / TP 8 | Random; 2,048 input, 256 output, concurrency 64, 128 prompts | 1,580 ms | 24.1 ms | 2,385 tokens/s |
| 8 × H200 | BF16 / TP 8 | Random; 2,048 input, 256 output, concurrency 64, 128 prompts | 1,054 ms | 70.8 ms | 1,044 tokens/s |
Both measurements use SGLang PR `#27944`, CUDA graphs, a flushed cache, and warm steady-state runs. Peak GPU memory, host memory, cold-start time, and a production concurrency recommendation were not published. Benchmark your prompt lengths, output lengths, multimodal ratio, and concurrency before capacity planning.
## Weights, cache, and offline deployment
The runtime downloads Hugging Face files to `/root/.cache/huggingface` in the container, mapped to `~/.cache/huggingface` on the host by the Quickstart command.
To prepare an offline copy, install the [Hugging Face CLI](https://huggingface.co/docs/huggingface_hub/guides/cli) on a connected machine and download the pinned revision:
```bash theme={null}
hf download MiniMaxAI/MiniMax-M3-MXFP8 \
--revision c5454eb03678d8710e54a4e0fc681b9f3b4a3dba \
--local-dir /data/models/MiniMax-M3-MXFP8
```
Copy that directory to the deployment host, then start without Hub access:
```bash theme={null}
docker run --rm --name minimax-m3 \
--gpus all \
--shm-size 32g \
--ipc=host \
-p 127.0.0.1:30000:30000 \
-e HF_HUB_OFFLINE=1 \
-v /data/models/MiniMax-M3-MXFP8:/models/MiniMax-M3-MXFP8:ro \
lmsysorg/sglang@sha256:de63ac56df5d7b064451e21147eaab89634a02332d830ca8c01cb8c033b3a78f \
sglang serve \
--trust-remote-code \
--model-path /models/MiniMax-M3-MXFP8 \
--reasoning-parser auto \
--tool-call-parser auto \
--tp 8 \
--attention-backend fa4 \
--mm-attention-backend flashinfer_cudnn \
--moe-runner-backend deep_gemm \
--chunked-prefill-size 8192 \
--mem-fraction-static 0.65 \
--host 0.0.0.0 \
--port 30000
```
Official ModelScope mirrors are available at [`MiniMax/MiniMax-M3-MXFP8`](https://modelscope.cn/models/MiniMax/MiniMax-M3-MXFP8) and [`MiniMax/MiniMax-M3`](https://modelscope.cn/models/MiniMax/MiniMax-M3). The ModelScope pages do not identify snapshots with the Hugging Face commit used by this baseline, so revalidate the files before substituting a mirror in a revision-pinned deployment.
If a download is interrupted, rerun `hf download` with the same revision and directory; the client resumes from its cache. If loading reports missing or corrupt shards, verify available disk space and redownload the affected revision before changing runtime parameters.
## Production and security
Keep the loopback binding for development. Before exposing the service to another host:
* Set an unpredictable SGLang `--api-key` and send it as `Authorization: Bearer `.
* Terminate TLS at a reverse proxy or trusted ingress, restrict source networks with a firewall or security group, and add request-rate and concurrency limits.
* Do not log API keys, full prompts, Base64 media, or model outputs unless your data policy explicitly permits it.
* Reject arbitrary remote media URLs at the gateway. Apply a domain allowlist, block private and link-local destinations, limit redirects, download time, byte size, image pixels, and video duration, and validate media types before forwarding a request.
* Monitor GPU memory, queue depth, request latency, error rates, process health, and disk utilization. A successful `/health` response alone is not a production readiness test.
SGLang accepts `--api-key`, but the current MiniMax-M3 Cookbook does not provide a complete remote-media security policy. Enforce those controls in your gateway and network layer.
## Troubleshooting
| Symptom | Likely cause | Shortest corrective action |
| :------------------------------------------------------------------------ | :---------------------------------------------------------------------------------------- | :------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| `CUDA out of memory` during startup or prefill | Insufficient activation headroom or an oversized prefill chunk | Reduce `--mem-fraction-static` below `0.65`, then reduce `--chunked-prefill-size` to `4096` or `2048`; this reduces KV capacity or prefill speed, so benchmark again |
| Weights fail to load or a shard is missing | Incomplete download, insufficient disk, or a different revision | Confirm the model path, free disk, and revision; rerun `hf download` with the pinned revision |
| `AttributeError: Module has no function 'plan'` during CUDA graph capture | Multiple TP ranks raced while JIT-compiling the MSA kernel | Run the SGLang Cookbook's single-process `msa_available()` gate once in the same image and cache, then restart the server |
| MSA is not active on B200 | Wrong image, unsupported GPU architecture, or missing `fa4` backend | Confirm the pinned image and `--attention-backend fa4`; follow the Cookbook gate check and expect `True` |
| NCCL initialization hangs or times out | Not all GPUs are visible, insufficient shared memory, or a topology/driver problem | Verify eight GPUs with `nvidia-smi`, preserve `--ipc=host` and `--shm-size 32g`, then enable `NCCL_DEBUG=INFO` and inspect the first failing rank |
| Raw `` or MiniMax tool-call tokens appear in `content` | The corresponding parser was not enabled or the runtime differs from the pinned image | Restore `--reasoning-parser auto` and `--tool-call-parser auto`, then restart with the pinned digest |
| `/health` fails after the process starts | The model is still downloading/loading, a worker exited, or port publication is incorrect | Inspect container logs, verify host disk and GPU state, and confirm `127.0.0.1:30000:30000` is published |
| A remote image request fails | The server cannot reach the URL, or the gateway blocks it | Use an allowlisted URL or a validated Base64 data URI; do not disable network protections globally |
## License and use restrictions
The BF16 and MXFP8 weights are released under the [MiniMax Community License](https://huggingface.co/MiniMaxAI/MiniMax-M3/blob/main/LICENSE). It includes attribution requirements, commercial notice or authorization requirements, and prohibited uses. Read the complete license and complete the applicable process before production or commercial use; this page is not legal advice.
## Resources
Review the upstream hardware matrix, configuration generator, benchmark conditions, and advanced tuning.
Review the official weights, model capabilities, inference parameters, and license.
Open the MXFP8 checkpoint used by the reference deployment.
Review the MSA kernel source and Blackwell requirements.