Skip to content

Supported models

The tables on this page are generated from Dew’s source at build time. Pretrained.load(source) reads a Hugging Face directory or Hub repository whose config.json names one of the model_type values below, and dew.pipeline(source) wraps the same load in a generation task. Pretrained.save writes trained weights back in the source’s own layout. Transformers loads the exported directory for the registered causal decoder families; a text-only DiffusionGemma export is refused before anything is written.

A port counts as supported when a test loads the same weights into Dew and into the reference implementation and compares the outputs in float32. For most families that test runs on a small fixture with the release’s own configuration and tensor shapes. The “Checked with” column says when a family was also compared at full size against a released checkpoint. Most families have not been run at full size; memory use and throughput at full size are unmeasured for them.

A checkpoint whose type is not listed can still load. A type that computes a registered family’s block is verified against transformers at load time. Any other causal LM that Transformers can build loads through Transformers’ own PyTorch forward, lowered to JAX by torchax, its attention Dew’s (packing included) wherever its layers ask Transformers’ attention interface for it, generating through Dew’s decode cache where its attention is Dew’s, with no sharding rules, and fallback="torchax" takes that route for any checkpoint. GGUF files, pytorch_model.bin repositories and mamba_ssm checkpoints load too. Train language models explains each of these paths and what it costs.

These families were asked for and are not in the tables above. Each one is tracked until it loads and a test compares it with its reference:

FamilyWhat is missing
MiniMax M3The shared expert, MTP and vision-language half, and a reader.
LLaDA MoE (GSAI LLaDA-MoE-v2-30B-A3B, inclusionAI LLaDA-MoE-7B-A1B) and LLaDA 2.xThe expert layers and routers in the LLaDA reader, which reads the dense 8B model.
Apple DiffuCoderA pinned load compared with the released model. The Dream reader targets Dream-v0.
Kimi K2.5 and K3 visionThe vision towers and projectors. Both load as text decoders.
MiniMax M2.5, M2.7 and H3Their configurations, to check whether the M2 reader covers them.
Familymodel_typeArchitectureChecked with
GPT-2gpt2GPT2LMHeadModelA fixture shaped like the release
GPT-NeoXgpt_neoxGPTNeoXForCausalLMA fixture shaped like the release
OPToptOPTForCausalLMA fixture shaped like the release
GPT-Neogpt_neoGPTNeoForCausalLMA fixture shaped like the release
GPT-1openai-gptOpenAIGPTLMHeadModelA fixture shaped like the release
PhiphiPhiForCausalLMA fixture shaped like the release
Phi-3phi3Phi3ForCausalLMA fixture shaped like the release
FalconfalconFalconForCausalLMA fixture shaped like the release
GPT-JgptjGPTJForCausalLMA fixture shaped like the release
GPTBigCodegpt_bigcodeGPTBigCodeForCausalLMA fixture shaped like the release
StarCoder2starcoder2Starcoder2ForCausalLMA fixture shaped like the release
LlamallamaLlamaForCausalLMSmolLM2-135M, logits against transformers; its GGUF Q8_0 and Q4_K_M files
MistralmistralMistralForCausalLMA fixture shaped like the release
MinistralministralMinistralForCausalLMA fixture shaped like the release
Qwen 2qwen2Qwen2ForCausalLMA fixture shaped like the release
Qwen 3qwen3Qwen3ForCausalLMQwen3-0.6B, logits against transformers, on one GPU, and streamed onto an 8-device CPU mesh
Ouro (looped)ouroOuroForCausalLMA fixture shaped like the release
Dory (looped nGPT)doryDoryForCausalLMA fixture shaped like the release
Qwen 3.5, textqwen3_5_textQwen3_5ForCausalLMA fixture shaped like the release
K2 Horizonk2_horizonK2HorizonForCausalLMA fixture shaped like the release
Muse Glimmer, textmuse_glimmer_textMuseGlimmerTextModelA fixture shaped like the release
BLOOMbloomBloomForCausalLMA fixture shaped like the release
GemmagemmaGemmaForCausalLMA fixture shaped like the release
Gemma 2gemma2Gemma2ForCausalLMA fixture shaped like the release
Gemma 3, textgemma3_textGemma3ForCausalLMA fixture shaped like the release
Gemma 3n, textgemma3n_textGemma3nForCausalLMA fixture shaped like the release
Gemma 4, textgemma4_textGemma4ForCausalLMA fixture shaped like the release
Gemma 4 Unified, textgemma4_unified_textGemma4UnifiedForCausalLMA fixture shaped like the release
OLMo 3olmo3Olmo3ForCausalLMA fixture shaped like the release
Familymodel_typeArchitectureChecked with
MiMo-V2mimo_v2MiMoV2ForCausalLMA fixture shaped like the release
MixtralmixtralMixtralForCausalLMA fixture shaped like the release
MiniMax-M2minimax_m2MiniMaxM2ForCausalLMA fixture shaped like the release
MiMo-V2-Flashmimo_v2_flashMiMoV2FlashForCausalLMA fixture shaped like the release
Granite MoEgranitemoeGraniteMoeForCausalLMA fixture shaped like the release
Qwen3-MoEqwen3_moeQwen3MoeForCausalLMA fixture shaped like the release
Qwen 2 MoEqwen2_moeQwen2MoeForCausalLMA fixture shaped like the release
Qwen 3.5 MoE, textqwen3_5_moe_textQwen3_5MoeForCausalLMA fixture shaped like the release
Qwen3-VL-MoE, textqwen3_vl_moe_textQwen3MoeForCausalLMA fixture shaped like the release
gpt-ossgpt_ossGptOssForCausalLMA fixture shaped like the release
Llama 4, textllama4_textLlama4ForCausalLMA fixture shaped like the release
GLM 4 MoEglm4_moeGlm4MoeForCausalLMA fixture shaped like the release
GLM-5.3glm_moe_dsaGlmMoeDsaForCausalLMA fixture shaped like the release
DeepSeek V2deepseek_v2DeepseekV2ForCausalLMA fixture shaped like the release
DeepSeek V3deepseek_v3DeepseekV3ForCausalLMA fixture shaped like the release
DeepSeek V3.2deepseek_v32DeepseekV32ForCausalLMA fixture shaped like the release
DeepSeek V4deepseek_v4DeepseekV4ForCausalLMA fixture shaped like the release
DeepSeek V4.1deepseek_v41DeepseekV41ForCausalLMA fixture shaped like the release
Kimi K2kimi_k2DeepseekV3ForCausalLMA fixture shaped like the release
Kimi K2.5, textkimi_k25Kimi_K25ForConditionalGenerationA fixture shaped like the release
Familymodel_typeArchitectureChecked with
Qwen3-Nextqwen3_nextQwen3NextForCausalLMA fixture shaped like the release
Nemotron-Hnemotron_hNemotronHForCausalLMA fixture shaped like the release
GLM-5.3-Flash, textglm5_next_textGlm5NextTextForCausalLMA fixture shaped like the release
Kimi Linearkimi_linearKimiLinearForCausalLMA fixture shaped like the release
LFM3lfm3Lfm3ForCausalLMA fixture shaped like the release
Kimi K3, textkimi_k3KimiK3ForConditionalGenerationA fixture shaped like the release
MiniMax-Text-01 and MiniMax-M1minimax, minimax_text_01, minimax_m1MiniMaxForCausalLMA fixture shaped like the release
Qwen4-Exp (Qwen3.8-Flash-Next), textqwen4_exp_textQwen4ExpForCausalLMA fixture shaped like the release
MambamambaMambaForCausalLMA fixture shaped like the release
Mamba-2mamba2Mamba2ForCausalLMstate-spaces/mamba2-130m against AntonV/mamba2-130m-hf
Familymodel_typeArchitectureChecked with
LLaDAlladaLLaDAModelLMA fixture shaped like the release
LLaDA-MoE v2llada_moe_v2LLaDAMoEv2ModelLMA fixture shaped like the release
LLaDA2 MoEllada2_moeLLaDA2MoeModelLMA fixture shaped like the release
Dreamdream, DreamDreamModelA fixture shaped like the release
Diffusion Gemmadiffusion_gemma (text config diffusion_gemma_text)DiffusionGemmaForBlockDiffusionA fixture shaped like the release
Familymodel_typeArchitectureChecked with
ModernBERTmodernbertModernBertForMaskedLMA fixture shaped like the release

These load the whole checkpoint: the text decoder listed above, the media towers and the checkpoint’s own processor, but where the media column says the language model alone, whose towers have no port yet.

Familymodel_typeMedia
Gemma 3gemma3Images
Llama 4llama4Images
Gemma 4gemma4Images, video, audio
Gemma 4 Unifiedgemma4_unifiedImages, video, audio
Qwen 3.5qwen3_5Images, video
Qwen 3.5 MoEqwen3_5_moeImages, video
Qwen4-Exp (Qwen3.8-Flash-Next)qwen4_expImages, video
Gemma 3ngemma3nImages, audio
Muse Glimmermuse_glimmerImages
LFM3-VLlfm3_vlImages, video
Qwen2-VLqwen2_vlNone yet: its language model alone (text_only)
Qwen2.5-VLqwen2_5_vlNone yet: its language model alone (text_only)
Qwen3-VLqwen3_vlNone yet: its language model alone (text_only)
Qwen3-VL-MoEqwen3_vl_moeNone yet: its language model alone (text_only)
DeepSeek V4.1deepseek_v41Images

Pretrained.load and dew.pipeline read a diffusers pipeline directory by the class it names in model_index.json.

FamilyPipeline classTask
Stable DiffusionStableDiffusionPipelineText to image
Stable DiffusionStableDiffusionImg2ImgPipelineImage to image
Stable DiffusionStableDiffusionInpaintPipelineInpainting
Stable Diffusion XLStableDiffusionXLPipelineText to image
Stable Diffusion XLStableDiffusionXLImg2ImgPipelineImage to image, refiner
Stable Diffusion XLStableDiffusionXLInpaintPipelineInpainting
Stable Diffusion 3StableDiffusion3PipelineText to image
FluxFluxPipelineText to image
Qwen-Image 2.1QwenImage21PipelineText to image
FLUX.2 [dev]Flux2PipelineText to image
FLUX.2 [klein]Flux2KleinPipelineText to image
Z-ImageZImagePipelineText to image
Wan 2.1WanPipelineText to video
Stable Diffusion, Flax weightsFlaxStableDiffusionPipelineText to image
Stable Diffusion, Flax weightsFlaxStableDiffusionImg2ImgPipelineImage to image
Stable Diffusion, Flax weightsFlaxStableDiffusionInpaintPipelineInpainting
Stable Diffusion XL, Flax weightsFlaxStableDiffusionXLPipelineText to image

Code builds each one from its class; a run record names it by its registered name. A class the API reference documents links to its entry.

Registered nameClass
causal_transformerCausalTransformer
diffusion_gemmaDiffusionGemma
edm2_unetEDM2UNet
flux2_transformerFlux2Transformer
flux_transformerFluxTransformer
hierarchical_mmditHierarchicalMMDiT
hybrid_ditHybridSSMAttentionDiT
jepa_encoderJepaEncoder
jepa_predictorJepaPredictor
jepa_video_encoderJepaVideoEncoder
multimodal_transformerMultimodalTransformer
qwen_image_transformerQwenImageTransformer
sd3_transformerSD3Transformer
simple_ditSimpleDiT
simple_mmditSimpleMMDiT
simple_uditSimpleUDiT
unetUnet
unet_2d_conditionUNet2DCondition
unet_3dUNet3D
uvitUViT
video_ditVideoDiT
wan_transformerWanTransformer
z_image_transformerZImageTransformer