MiniMaxAI/MiniMax-M3-MXFP8
Original427.0Bparameters
Fixture31.0Mparameters
Layer stacks2compared
Findings0to review
✓ Same architecture, smaller
Every module class, layer kind and checked property of the original is in the fixture.
Layer stacks
One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.
model.vision_tower.layers
Original · 32
Fixture · 2
| Kind | Original | Fixture | |
|---|---|---|---|
| MiniMaxM3VLVisionEncoderLayer | 32 | 2 | ✓ kept |
model.language_model.layers
Original · 60
Fixture · 2
| Kind | Original | Fixture | |
|---|---|---|---|
| no MiniMaxM3VLExperts | 3 | 1 | ✓ kept |
| MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock … | 57 | 1 | ✓ kept |
Where the parameters are
Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.
Original
Fixture
| Component | Original | Share | Fixture | Share |
|---|---|---|---|---|
| model.language_model.layers | 423.7B | 99.2% | 2.5M | 8.2% |
| model.language_model | 1.2B | 0.3% | 12.8M | 41.2% |
| lm_head | 1.2B | 0.3% | 12.8M | 41.2% |
| model.vision_tower.layers | 629.7M | 0.1% | 66.9K | 0.2% |
| model.multi_modal_projector | 234.4M | 0.1% | 2.8M | 8.9% |
| model.vision_tower | 1.5M | 0.0% | 75.4K | 0.2% |
Preserved properties
| Config | Property | Original | Fixture | |
|---|---|---|---|---|
| model | model_type | minimax_m3_vl | minimax_m3_vl | ✓ kept |
| model | tie_word_embeddings | False | False | ✓ kept |
| vision_config | RoPE type | axial | axial | ✓ kept |
| vision_config | model_type | minimax_m3_vl_vision | minimax_m3_vl_vision | ✓ kept |
| vision_config | vocab_size | 32000 | 32000 | ✓ kept |
| vision_config | hidden_act | gelu | gelu | ✓ kept |
| vision_config | image_size | 2016 | 2016 | ✓ kept |
| vision_config | patch_size | 14 | 14 | ✓ kept |
| vision_config | num_channels | 3 | 3 | ✓ kept |
| text_config | attention style | GQA (64 heads, 4 KV) | GQA (4 heads, 2 KV) | ✓ kept |
| text_config | MoE routing | top-4 of 128 experts | top-2 of 4 experts | ✓ kept |
| text_config | RoPE type | default | default | ✓ kept |
| text_config | model_type | minimax_m3_vl_text | minimax_m3_vl_text | ✓ kept |
| text_config | vocab_size | 200064 | 200064 | ✓ kept |
| text_config | tie_word_embeddings | False | False | ✓ kept |
| text_config | hidden_act | silu | silu | ✓ kept |
| text_config | max_position_embeddings | 1048576 | 1048576 | ✓ kept |
Module classes
Instances of each class in the built model (prefix MiniMaxM3Sparse omitted).
| Class | Original | Fixture |
|---|---|---|
| Linear | 671 | 31 |
| MiniMaxM3VLRMSNorm | 355 | 11 |
| LayerNorm | 65 | 5 |
| MiniMaxM3VLAttention | 60 | 2 |
| MiniMaxM3VLDecoderLayer | 60 | 2 |
| MiniMaxM3VLDenseMLP | 60 | 2 |
| MiniMaxM3VLExperts | 57 | 1 |
| MiniMaxM3VLIndexer | 57 | 1 |
| MiniMaxM3VLSparseMoeBlock | 57 | 1 |
| MiniMaxM3VLTopKRouter | 57 | 1 |
| GELUActivation | 34 | 4 |
| MiniMaxM3VLVisionAttention | 32 | 2 |
| MiniMaxM3VLVisionEncoderLayer | 32 | 2 |
| MiniMaxM3VLVisionMLP | 32 | 2 |
| ModuleList | 2 | 2 |
| Conv3d | 1 | 1 |
| Embedding | 1 | 1 |
| ForConditionalGeneration | 1 | 1 |
| MiniMaxM3VLModel | 1 | 1 |
| MiniMaxM3VLMultiModalProjector | 1 | 1 |
| MiniMaxM3VLRotaryEmbedding | 1 | 1 |
| MiniMaxM3VLTextModel | 1 | 1 |
| MiniMaxM3VLVisionEmbeddings | 1 | 1 |
| MiniMaxM3VLVisionModel | 1 | 1 |
| MiniMaxM3VLVisionRotaryEmbedding | 1 | 1 |
Config changes
21 fields changed, 91 unchanged.
| Field | Original | Fixture |
|---|---|---|
| merged_hidden_size | 24576 | 256 |
| quantization_config | {"quant_method": "mxfp8", "activation_scheme": "dynamic", "weight_block_size"… | — |
| text_config.head_dim | 128 | 16 |
| text_config.hidden_size | 6144 | 64 |
| text_config.index_block_size | 128 | 8 |
| text_config.index_n_heads | 4 | 2 |
| text_config.index_topk_blocks | 16 | 4 |
| text_config.intermediate_size | 3072 | 128 |
| text_config.layer_types | 60 entries: 57× minimax_m3_sparse, 3× full_attention | ["full_attention", "minimax_m3_sparse"] |
| text_config.mlp_layer_types | 60 entries: 57× sparse, 3× dense | ["dense", "sparse"] |
| text_config.num_attention_heads | 64 | 4 |
| text_config.num_experts_per_tok | 4 | 2 |
| text_config.num_hidden_layers | 60 | 2 |
| text_config.num_key_value_heads | 4 | 2 |
| text_config.num_local_experts | 128 | 4 |
| text_config.shared_intermediate_size | 3072 | 128 |
| vision_config.hidden_size | 1280 | 64 |
| vision_config.intermediate_size | 5120 | 128 |
| vision_config.num_attention_heads | 16 | 4 |
| vision_config.num_hidden_layers | 32 | 2 |
| vision_config.projection_dim | 6144 | 64 |