Supported architectures¶

The table below lists every Hugging Face architecture that modelbuilder.builder.create_model() can convert, together with the builder class that handles it. It is generated automatically at documentation build time by parsing the dispatch chain of create_model (see modelbuilder.architectures.list_supported_architectures()), so it is always in sync with the code.

The architecture name is the value of config.architectures[0] in the Hugging Face config.json. To convert a model of a given family, point -m or -i at a checkpoint whose config declares that architecture.

49 architectures are currently supported.

Architecture	Builder
`ChatGLMForConditionalGeneration`	`modelbuilder.builders.chatglm.ChatGLMModel`
`ChatGLMModel`	`modelbuilder.builders.chatglm.ChatGLMModel`
`DeepseekV3ForCausalLM`	`modelbuilder.builders.deepseek.DeepSeekV3Model`
`DeepseekV4ForCausalLM`	`modelbuilder.builders.deepseek.DeepSeekV3Model`
`Ernie4_5ForCausalLM`	`modelbuilder.builders.ernie.ErnieModel`
`Gemma2ForCausalLM`	`modelbuilder.builders.gemma.Gemma2Model`
`Gemma3ForCausalLM`	`modelbuilder.builders.gemma.Gemma3Model`
`Gemma3ForConditionalGeneration`	`modelbuilder.builders.gemma.Gemma3Model`
`Gemma4ForCausalLM`	`modelbuilder.builders.gemma.Gemma4Model`
`Gemma4ForConditionalGeneration`	`modelbuilder.builders.gemma.Gemma4Model`
`GemmaForCausalLM`	`modelbuilder.builders.gemma.GemmaModel`
`GPT2LMHeadModel`	`modelbuilder.builders.gpt2.GPT2Model`
`GptOssForCausalLM`	`modelbuilder.builders.gptoss.GPTOSSModel`
`GraniteForCausalLM`	`modelbuilder.builders.granite.GraniteModel`
`HunYuanDenseV1ForCausalLM`	`modelbuilder.builders.hunyuan.HunyuanDenseV1Model`
`InternLM2ForCausalLM`	`modelbuilder.builders.internlm.InternLM2Model`
`Lfm2ForCausalLM`	`modelbuilder.builders.lfm2.LFM2Model`
`LlamaForCausalLM`	`modelbuilder.builders.llama.LlamaModel`
`MageFlowForCausalLM`	`modelbuilder.builders.mageflow.MageFlowModel`
`Ministral3ForCausalLM`	`modelbuilder.builders.mistral.Ministral3TextModel`
`Mistral3ForConditionalGeneration`	`modelbuilder.builders.mistral.Ministral3ConditionalGenerationModel`
`MistralForCausalLM`	`modelbuilder.builders.mistral.MistralModel`
`MistralNeMoForCausalLM`	`modelbuilder.builders.mistral.MistralNeMoModel`
`MixtralForCausalLM`	`modelbuilder.builders.mistral.MixtralModel`
`NemotronForCausalLM`	`modelbuilder.builders.nemotron.NemotronModel`
`NemotronHForCausalLM`	`modelbuilder.builders.nemotron.NemotronHModel`
`Olmo2ForCausalLM`	`modelbuilder.builders.olmo.OLMo2Model`
`Olmo3ForCausalLM`	`modelbuilder.builders.olmo.OLMo3Model`
`OlmoForCausalLM`	`modelbuilder.builders.olmo.OLMoModel`
`Phi3ForCausalLM`	`modelbuilder.builders.phi.Phi3MiniModel`
`Phi3SmallForCausalLM`	`modelbuilder.builders.phi.Phi3SmallModel`
`Phi3VForCausalLM`	`modelbuilder.builders.phi.Phi3VModel`
`Phi4MMForCausalLM`	`modelbuilder.builders.phi.Phi4MMModel`
`Phi4MultimodalForCausalLM`	`modelbuilder.builders.phi.Phi4MultimodalConditionalGenerationModel`
`PhiForCausalLM`	`modelbuilder.builders.phi.PhiModel`
`PhiMoEForCausalLM`	`modelbuilder.builders.phi.Phi3MoELongRoPEModel`
`Qwen2_5_VLForConditionalGeneration`	`modelbuilder.builders.qwen.Qwen25VLTextModel`
`Qwen2_5OmniForConditionalGeneration`	`modelbuilder.builders.qwen.Qwen25OmniConditionalGenerationModel`
`Qwen2_5OmniThinkerForConditionalGeneration`	`modelbuilder.builders.qwen.Qwen25OmniConditionalGenerationModel`
`Qwen2ForCausalLM`	`modelbuilder.builders.qwen.QwenModel`
`Qwen3_5ForCausalLM`	`modelbuilder.builders.qwen.Qwen35CausalLMModel`
`Qwen3_5ForConditionalGeneration`	`modelbuilder.builders.qwen.Qwen35TextModel`
`Qwen3_5MoeForCausalLM`	`modelbuilder.builders.qwen.Qwen35MoeCausalLMModel`
`Qwen3_5MoeForConditionalGeneration`	`modelbuilder.builders.qwen.Qwen35MoeTextModel`
`Qwen3ForCausalLM`	`modelbuilder.builders.qwen.Qwen3Model`
`Qwen3VLForConditionalGeneration`	`modelbuilder.builders.qwen.Qwen3VLTextModel`
`SmolLM3ForCausalLM`	`modelbuilder.builders.smollm.SmolLM3Model`
`VideoChatFlashQwenForCausalLM`	`modelbuilder.builders.qwen.VideoChatFlashQwenModel`
`WhisperForConditionalGeneration`	`modelbuilder.builders.whisper.WhisperModel`

Adding a new architecture¶

Support for a new family is added by:

implementing a Model subclass in modelbuilder/builders/ (usually subclassing an existing builder and overriding the few pieces that differ);
adding an elif config.architectures[0] == "<Name>ForCausalLM": branch to create_model() that instantiates the builder;
adding a fast test under tests/fast that converts a tiny random-weight model and checks the discrepancies.

Once the branch is added, this page picks up the new architecture automatically.