THE SOURCE IN CONTEXT
The pinned MiniMax-M3 recipe defaults to MiniMaxAI/MiniMax-M3 (bf16). Its effective default-variant minimum is vLLM 0.24.0. Select the exact variant, hardware override and topology before treating it as a launch plan. No deployment was run by this site.
Which artifact and engine does this recipe select?
The recipe identifies MiniMaxAI/MiniMax-M3 as its model family entry and selects MiniMaxAI/MiniMax-M3 by default. The default precision label is bf16; model-wide and variant-specific runtime requirements are kept separately below. The document revision fixes what the recipe said, while the weight revision and executed runtime revision remain separate fields. A model-family title alone cannot establish either artifact identity.
Source: vllm-project · pinned document
What needs special attention for this model?
The default is BF16 and is much larger than the three-Spark NVFP4 operator setup. The recipe’s 427B total / 26B active labels differ from the official card’s approximately 428B / 23B. Neither source is silently corrected here. The block-size setting, attention backend and DSpark choice travel with the selected precision and topology.
Source: vllm-project · pinned document
How should the settings be applied?
Choose one artifact variant and one hardware profile from the original recipe, then resolve its shared, variant, hardware, strategy and optional-feature settings together. The base arguments listed here are only one layer and are deliberately not presented as a complete runnable command. Check any nightly-image requirement and the model’s parsers before loading weights. A context declaration is a model limit, while batched-token and sequence settings control a serving workload.
Source: vllm-project · pinned document
What does the memory number mean?
The GB figure is the recipe’s declared planning minimum for the selected variant. It is not a site measurement, a universal peak allocation or a promise of maximum context. KV cache, auxiliary/draft models, image encoders, temporary workspace and active requests still require memory. Keep GB and GiB distinct, and use the actual artifact size and runtime allocation for a machine-specific budget.
Source: vllm-project · pinned document
What is needed to reproduce a deployment?
Save the model and conversion revisions, weight hashes, tokenizer/template, container or runtime digest, device count/topology and complete resolved launch. First check one short public request. For a subsequent measurement, fix input and output counts, concurrency, cache state, warmup, sampling, reasoning effort, repetitions and failures. The declarations on this page are source configuration facts; no throughput or quality outcome follows from them.
Source: vllm-project · pinned document
What does the source specify?
These declarations belong to the cited document. A planning minimum, model limit and measured runtime allocation are different quantities.
| Architecture | moe |
|---|---|
| Total parameters as counted by recipe | 427B |
| Active parameters as counted by recipe | 26B |
| Native context declaration (tokens) | 1048576 |
| Model-wide minimum vLLM | 0.24.0 |
| Effective default-variant minimum vLLM | 0.24.0 |
| Default artifact | MiniMaxAI/MiniMax-M3 |
| Default precision label | bf16 |
| Default planning minimum (GB; recipe declaration) | 1025 |
| Default variant hardware allowlist | Not specified |
| Shared base arguments (incomplete launch) | ["--block-size", "128"] |
| Declared compatible strategies | ["single_node_tp", "single_node_tep", "single_node_dep", "multi_node_tp", "multi_node_tp_pp", "multi_node_tep", "multi_node_dep", "pd_cluster"] |
| Optional feature names | ["tool_calling", "reasoning", "spec_decoding", "text_only", "thinking_always_on", "encoder_parallel"] |
| Features selected by default | ["spec_decoding", "text_only", "thinking_always_on", "encoder_parallel"] |
AI-assisted source reading and original editorial synthesis; no independent model execution.