This page is generated from the registries in Dew’s source at build time, so it lists what the code loads today. load_pretrained(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, so the directory loads again in transformers.
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. I have not run most families at full size, so check memory and throughput yourself before a large run.
A checkpoint whose type is not listed can still load. A type that computes the Llama block is verified against transformers at load time, and fallback="torchax" runs any transformers causal LM through its own PyTorch code. GGUF files, pytorch_model.bin repositories and mamba_ssm checkpoints load too. Train language models explains each of these paths and what it costs.
| Family | model_type | Architecture | Checked with |
|---|
| Llama | llama | LlamaForCausalLM | SmolLM2-135M, logits against transformers; its GGUF Q8_0 and Q4_K_M files |
| Mistral | mistral | MistralForCausalLM | A fixture shaped like the release |
| Ministral | ministral | MinistralForCausalLM | A fixture shaped like the release |
| Qwen 2 | qwen2 | Qwen2ForCausalLM | A fixture shaped like the release |
| Qwen 3 | qwen3 | Qwen3ForCausalLM | Qwen3-0.6B, logits against transformers, on one GPU and streamed onto two |
| Qwen 3.5, text | qwen3_5_text | Qwen3_5ForCausalLM | A fixture shaped like the release |
| Gemma | gemma | GemmaForCausalLM | A fixture shaped like the release |
| Gemma 2 | gemma2 | Gemma2ForCausalLM | A fixture shaped like the release |
| Gemma 3, text | gemma3_text | Gemma3ForCausalLM | A fixture shaped like the release |
| Gemma 3n, text | gemma3n_text | Gemma3nForCausalLM | A fixture shaped like the release |
| Gemma 4, text | gemma4_text | Gemma4ForCausalLM | A fixture shaped like the release |
| OLMo 3 | olmo3 | Olmo3ForCausalLM | A fixture shaped like the release |
| Family | model_type | Architecture | Checked with |
|---|
| Mixtral | mixtral | MixtralForCausalLM | A fixture shaped like the release |
| Qwen3-MoE | qwen3_moe | Qwen3MoeForCausalLM | A fixture shaped like the release |
| Qwen 3.5 MoE, text | qwen3_5_moe_text | Qwen3_5MoeForCausalLM | A fixture shaped like the release |
| gpt-oss | gpt_oss | GptOssForCausalLM | A fixture shaped like the release |
| Llama 4, text | llama4_text | Llama4ForCausalLM | A fixture shaped like the release |
| GLM 4 MoE | glm4_moe | Glm4MoeForCausalLM | A fixture shaped like the release |
| GLM MoE with sparse attention | glm_moe_dsa | GlmMoeDsaForCausalLM | A fixture shaped like the release |
| DeepSeek V2 | deepseek_v2 | DeepseekV2ForCausalLM | A fixture shaped like the release |
| DeepSeek V3 | deepseek_v3 | DeepseekV3ForCausalLM | A fixture shaped like the release |
| DeepSeek V3.2 | deepseek_v32 | DeepseekV32ForCausalLM | A fixture shaped like the release |
| DeepSeek V4 | deepseek_v4 | DeepseekV4ForCausalLM | A fixture shaped like the release |
| Kimi K2 | kimi_k2 | DeepseekV3ForCausalLM | A fixture shaped like the release |
| Kimi K2.5, text | kimi_k25 | Kimi_K25ForConditionalGeneration | A fixture shaped like the release |
| Family | model_type | Architecture | Checked with |
|---|
| Qwen3-Next | qwen3_next | Qwen3NextForCausalLM | A fixture shaped like the release |
| GLM 5 Next, text | glm5_next_text | Glm5NextTextForCausalLM | A fixture shaped like the release |
| Kimi Linear | kimi_linear | KimiLinearForCausalLM | A fixture shaped like the release |
| Kimi K3, text | kimi_k3 | KimiK3ForConditionalGeneration | A fixture shaped like the release |
| Mamba-2 | mamba2 | Mamba2ForCausalLM | state-spaces/mamba2-130m against AntonV/mamba2-130m-hf |
| Family | model_type | Architecture | Checked with |
|---|
| LLaDA | llada | LLaDAModelLM | A fixture shaped like the release |
| Dream | dream, Dream | DreamModel | A fixture shaped like the release |
| Diffusion Gemma | diffusion_gemma_text | DiffusionGemmaForBlockDiffusion | A fixture shaped like the release |
These load the whole checkpoint: the text decoder listed above, the media towers and the checkpoint’s own processor.
| Family | model_type | Media |
|---|
| Gemma 3 | gemma3 | Images |
| Llama 4 | llama4 | Images |
| Gemma 4 | gemma4 | Images, video, audio |
| Qwen 3.5 | qwen3_5 | Images, video |
| Gemma 3n | gemma3n | Images, audio |
load_pretrained and dew.pipeline read a diffusers pipeline directory by the class it names in model_index.json.
| Family | Pipeline class | Task |
|---|
| Stable Diffusion | StableDiffusionPipeline | Text to image |
| Stable Diffusion | StableDiffusionImg2ImgPipeline | Image to image |
| Stable Diffusion | StableDiffusionInpaintPipeline | Inpainting |
| Stable Diffusion XL | StableDiffusionXLPipeline | Text to image |
| Stable Diffusion XL | StableDiffusionXLImg2ImgPipeline | Image to image, refiner |
| Stable Diffusion XL | StableDiffusionXLInpaintPipeline | Inpainting |
| Stable Diffusion 3 | StableDiffusion3Pipeline | Text to image |
| Flux | FluxPipeline | Text to image |
| Qwen-Image 2.1 | QwenImage21Pipeline | Text to image |
| Stable Diffusion, Flax weights | FlaxStableDiffusionPipeline | Text to image |
| Stable Diffusion, Flax weights | FlaxStableDiffusionImg2ImgPipeline | Image to image |
| Stable Diffusion, Flax weights | FlaxStableDiffusionInpaintPipeline | Inpainting |
| Stable Diffusion XL, Flax weights | FlaxStableDiffusionXLPipeline | Text to image |
models.build(name, ...) builds these by their registered name. Each links to its API entry.
