For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).

Python module

max.pipelines.architectures

Built-in model architectures for MAX pipelines.

Importing this package registers each architecture with PipelineRegistry; the registry imports an architecture’s module lazily, the first time that architecture is requested.

MAX includes built-in support for a wide range of model architectures. Each architecture module registers a SupportedArchitecture instance that tells the pipeline system how to load, configure, and execute a particular model family.

Text generation

deepseekV2	DeepSeek-V2 mixture-of-experts architecture for text generation.
deepseekV3	DeepSeek-V3 mixture-of-experts architecture for text generation.
deepseekV3_2	DeepSeek-V3.2 mixture-of-experts architecture for text generation.
deepseekV3_nextn	DeepSeek-V3 NextN multi-token prediction draft model for speculative decoding.
deepseekV4	DeepSeek-V4-Flash mixture-of-experts architecture for text generation.
dflash2_qwen3_5	DFlash2 draft model for Qwen3.5-family targets.
dflash_llama3	DFlash draft model for Llama3-family targets.
diffusion_gemma	DiffusionGemma block-diffusion architecture.
dspark_draft	Model-agnostic DSpark draft: block nn.Module + registry placeholder.
eagle3_deepseekV3	DeepseekV3 + Eagle3 speculator pipeline.
eagle_llama3	EAGLE speculative decoding draft model for Llama 3.
gemma3	Gemma 3 transformer architecture for text generation.
gemma3multimodal	Gemma 3 vision-language architecture for multimodal text generation.
gemma4	Gemma 4 vision-language architecture for multimodal text generation.
gemma4_assistant	Gemma 4 assistant draft architecture: a lightweight decoder that cross-attends to the Gemma 4 target's KV cache to draft speculative tokens.
glm5_1	GLM-5.1 (GlmMoeDsa) mixture-of-experts architecture for text generation.
glm5_next	GLM-5.3-Flash architecture: a multimodal MoE decoder with a vision tower, built on the DeepSeek V3.2 model.
gpt_oss	GPT-OSS mixture-of-experts architecture for text generation.
granite	Granite architecture, builds on llama3.
hy_v3	Tencent Hunyuan Hy3-preview (HYV3ForCausalLM).
idefics3	Idefics3 vision-language architecture for multimodal text generation.
inkling	Inkling architecture: a text-and-vision pipeline model with KV cache and SSM state support.
internvl	InternVL vision-language architecture for multimodal text generation.
kimik2_5	Kimi K2.5 mixture-of-experts architecture for text generation.
laguna	Laguna architecture: a decoder-only MoE model that extends the Llama 3 architecture.
lfm2	LFM2 architecture: a hybrid model combining full attention with short convolutions.
llama3	Llama 3 transformer architecture for text generation.
llama4	Llama 4 (text-only) transformer architecture for text generation.
mamba	Mamba state-space architecture for text generation.
mimo_v2	MiMo-V2.6-Flash (MiMoV2ForCausalLM).
minimax_m2	MiniMax-M2 architecture: a decoder-only MoE model that extends the Llama 3 architecture.
mistral	Mistral transformer architecture for text generation.
mistral3	Mistral 3 vision-language architecture for multimodal text generation.
muse_glimmer	Muse Glimmer vision-language architecture (ModuleV3 API).
nemotron_h	Nemotron-H architecture: a hybrid decoder interleaving Mamba-2 mixers, NoPE GQA attention, relu2 MLP blocks, and, in the Nemotron 3 MoE variants, mixture-of-experts blocks.
olmo	OLMo transformer architecture for text generation.
olmo2	OLMo 2 transformer architecture for text generation.
olmo3	OLMo 3 transformer architecture for text generation.
phi3	Phi-3 transformer architecture for text generation.
pixtral	Pixtral vision-language architecture for multimodal text generation.
qwen2	Qwen2 transformer architecture for text generation.
qwen2_5vl	Qwen2.5-VL vision-language architecture for multimodal text generation.
qwen3	Qwen3 transformer architecture for text generation.
qwen3_5	Qwen3.5 architecture: a hybrid model interleaving Gated DeltaNet linear attention with full attention, plus a vision encoder.
qwen3vl_moe	Qwen3-VL vision-language architecture for multimodal text generation.
step3p5	Step-3.5-Flash architecture, built on the Llama 3 model base.
unified_dflash2_qwen3_5	Qwen3.5 fused with a DFlash2 block drafter for speculative decoding.
unified_dflash_gemma4_31b	Unified DFlash speculative decoding for Gemma4 31B.
unified_dflash_kimi_k25	DFlash speculative decoding for Kimi K2.5 with unified graph compilation.
unified_dflash_llama3	DFlash speculative decoding for Llama3 with unified graph compilation.
unified_dflash_mimo_v2	MiMo-V2.6-Flash with its DFlash drafter: the fused speculative graph, and the base graph that writes the drafter's context.
unified_dspark_deepseekV4	DeepSeek-V4 speculative decoding with its in-checkpoint DSpark stages.
unified_dspark_gemma4_12b	Unified DSpark speculative decoding for Gemma4.
unified_dspark_gemma4_31b	Unified DSpark speculative decoding for Gemma4 31B.
unified_eagle_llama3	EAGLE speculative decoding draft model for Llama 3 with unified graph compilation.
unified_mtp_deepseekV3	DeepSeek-V3 multi-token prediction draft model for speculative decoding with unified graph compilation.
unified_mtp_gemma4	Gemma4 with MTP draft model for speculative decoding with unified graph compilation.
unified_mtp_glm5_2	GLM-5.2 (DeepSeek-V3.2 sparse) MTP draft model for unified speculative decoding.
unified_mtp_inkling	Inkling multi-token prediction for unified speculative decoding.
unified_mtp_qwen3_5	Qwen3.5 MTP head fused with its target for speculative decoding.

Embeddings

bert	BERT sentence transformer architecture for embeddings generation.
mpnet	MPNet sentence transformer architecture for embeddings generation.
qwen3_embedding	Qwen3 architecture for embeddings generation.

Image generation

flux2	FLUX.2 diffusion architecture for image generation.
ideogram4	Ideogram 4 flow-matching text-to-image architecture.
qwen_image	Qwen-Image diffusion architecture for image generation.
qwen_image_edit	Qwen-Image-Edit diffusion architecture for image editing.
wan	Wan diffusion architecture for video generation.

Audio generation

minimax_music3	Generates music from a text description and lyrics.