# MiniMax-M3: official model and deployment card

Original M3 weights and inference guidance, linking several local engines. This entry identifies the upstream model rather than a Spark-specific conversion.

- Type: Official model card
- Evidence: source_reported; site reproduced: false
- Model family: MiniMax
- Model: MiniMaxAI/MiniMax-M3
- Original author: MiniMaxAI
- Original project: https://huggingface.co/MiniMaxAI/MiniMax-M3
- Pinned document: https://huggingface.co/MiniMaxAI/MiniMax-M3/blob/f0e1c1e04d40177e4673a22097036854f536e9c0/README.md
- Source revision: f0e1c1e04d40177e4673a22097036854f536e9c0
- Source commit time (not experiment date): 2026-07-23T04:25:20.000Z
- Checked: 2026-10-17
- Experiment date: None
- Runtime topics: vLLM, SGLang, Transformers, KTransformers
- Hardware topics: 
- Reported setup: {"system": null, "node_count": null, "parallelism": null, "kv_dtype": null, "context_limit_tokens": null, "experiment_date": null}
- Performance figures: attributed selections with per-point conditions; no cross-source ranking

## Conditions

Unknown is null. Recipe options are not measured benchmark settings.

```json
{
  "conditions": {
    "model_revision": "f0e1c1e04d40177e4673a22097036854f536e9c0",
    "weight_hash": null,
    "runtime_execution_revision": null,
    "input_tokens": null,
    "output_tokens": null,
    "concurrency": null,
    "cache_state": null,
    "quantization": null,
    "speculative_decoding": null,
    "temperature": null,
    "top_p": null
  },
  "test_profiles": [],
  "recipe_parameters": null,
  "deployment_options": []
}
```

## Reading notes

- The publisher links model-specific engine documentation.
- Sampling recommendations are separate from a community benchmark’s actual controls.

## Limits

- Model terms and a converted artifact’s terms must both be checked.

## License declaration

{"declared_spdx": "other", "declared_name": "minimax-community", "scope": "Publisher model-card declaration; inspect the linked LICENSE and artifact-specific terms."}


## Editorial answer

Original M3 weights and inference guidance, linking several local engines. This entry identifies the upstream model rather than a Spark-specific conversion.

By Run AI Here editorial · 2026-10-17

AI-assisted source reading and original editorial synthesis; no independent model execution.

## Why does the model-size number differ between sources?

The official card gives approximately 428B total and 23B activated parameters, whereas the vLLM recipe lists 427B and 26B. The difference is unresolved in these references. We preserve both source declarations instead of averaging them or silently choosing one. The card advertises a native multimodal model with a 1M context window.

Sources: https://huggingface.co/MiniMaxAI/MiniMax-M3/blob/f0e1c1e04d40177e4673a22097036854f536e9c0/README.md#L34

## What does the sparse-attention speed claim mean?

MiniMax describes relative prefill and decode improvements over M2 at a one-million-token context. The card does not provide a matching workstation measurement for the reader’s hardware in that sentence. Treat it as a publisher comparison under its own conditions, and use the separately dated three-Spark operator report for an actual deployment account.

Sources: https://huggingface.co/MiniMaxAI/MiniMax-M3/blob/f0e1c1e04d40177e4673a22097036854f536e9c0/README.md#L38

## Which sampling settings are recommended?

The card recommends temperature 1.0 and top_p 0.95. Those recommendations differ from a greedy regression test at temperature zero. The three-Spark report contains both settings and reports different outcomes, so a reader should not combine those samples into a single accuracy or speed figure.

Sources: https://huggingface.co/MiniMaxAI/MiniMax-M3/blob/f0e1c1e04d40177e4673a22097036854f536e9c0/README.md#L90

## How should a local serving route be chosen?

The publisher’s local-deployment section links engine integrations. The companion recipe defaults to a large BF16 artifact, while the three-Spark deployment uses an NVIDIA NVFP4 conversion, a DSpark drafter and custom tensor-parallel patches. Those are different artifacts and environments. Capture the conversion and runtime revisions before using either as a reproduction plan.

Sources: https://huggingface.co/MiniMaxAI/MiniMax-M3/blob/f0e1c1e04d40177e4673a22097036854f536e9c0/README.md#L66

## Attributed configuration facts

```json
[
  {
    "name": "Card total / active parameters",
    "value": "approximately 428B / 23B",
    "classification": "publisher_declaration",
    "evidence_url": "https://huggingface.co/MiniMaxAI/MiniMax-M3/blob/f0e1c1e04d40177e4673a22097036854f536e9c0/README.md#L34",
    "source_document_sha256": "6148856dbf8be26dcf3f9b913be0252c38454e852108fdab01212fe965f3e327"
  },
  {
    "name": "Published context",
    "value": "1M tokens",
    "classification": "publisher_declaration",
    "evidence_url": "https://huggingface.co/MiniMaxAI/MiniMax-M3/blob/f0e1c1e04d40177e4673a22097036854f536e9c0/README.md#L34",
    "source_document_sha256": "6148856dbf8be26dcf3f9b913be0252c38454e852108fdab01212fe965f3e327"
  },
  {
    "name": "Recommended sampling",
    "value": "temperature 1.0; top_p 0.95",
    "classification": "publisher_declaration",
    "evidence_url": "https://huggingface.co/MiniMaxAI/MiniMax-M3/blob/f0e1c1e04d40177e4673a22097036854f536e9c0/README.md#L90",
    "source_document_sha256": "6148856dbf8be26dcf3f9b913be0252c38454e852108fdab01212fe965f3e327"
  }
]
```