Models
Mold supports image and video model families across a range of quality levels, hardware requirements, and creative workflows.
Community models from Civitai
The models documented here are not the limit. Open Models → Discover in Mold Studio and choose Civitai (or All) to browse and install compatible community checkpoints and LoRAs. See the Model Discovery Catalog for details.
Choosing a Model
| Need | Recommended | Why |
|---|---|---|
| Fast iterations | flux2-klein:q8 | 4 steps, ungated, Apache 2.0 |
| Best quality | flux-dev:q4 | 25 steps, excellent detail |
| Smallest checkpoint | flux2-klein:q4 | 2.6 GB transformer, 4 steps |
| Classic ecosystem | sd15:fp16 or dreamshaper-v8 | Huge model library, ControlNet |
| Fast + great | z-image-turbo:q8 | 9 steps, excellent quality |
| SDXL | sdxl-turbo:fp16 | 4 steps, 512x512 default (1024x1024 presets) |
| LTX Video | ltx-video-0.9.6-distilled:bf16 | Broad text-to-video default; LTX-2.x adds joint audio |
| Wan video | wan22-ti2v-5b:q8 | Text/image-to-video with broad Wan workflow support |
| Reference AV | minimax-h3-ref2va:comfy-pruned-int8 | Ordered image/video/audio references; SM89 CUDA build |
Choosing a Video Family
| Family | Start with | Best for | Important boundary |
|---|---|---|---|
| LTX Video | ltx-video-0.9.6-distilled:bf16 | Text-to-video plus joint audio-video on LTX-2.x | Legacy models have no generated audio; newer packs gated |
| Wan Video | wan22-ti2v-5b:q8 | Text/image-to-video, first/last frames, and sequences | CUDA-qualified; Metal/CPU correctness-only |
| MiniMax H3 | minimax-h3-fl2va:comfy-pruned-int8 | First-frame or ordered-reference audio-video | 42.482 GB base pull; runtime is build-scoped |
MiniMax H3's pull size is disk/download size, not peak VRAM. Its compact runtime is available on H3-enabled SM89 CUDA builds; Metal is shipped as an unqualified correctness-only route, and CPU is unsupported. Check the model row's runtime_available reason before downloading on another target.
Image VRAM Guide
These estimates include the transformer, text encoder(s), VAE, and ~2 GB activation headroom. The default column is sequential mode (drop-and-reload), which loads components one at a time. Eager mode keeps everything on GPU simultaneously for faster inference but needs more VRAM.
| Model | Variant | Default VRAM | Eager VRAM | Speed | Quality |
|---|---|---|---|---|---|
flux-schnell:q8 | Q8 | ~15 GB | ~25 GB | Fast, 4 steps | Good |
flux-dev:q4 | Q4 | ~10 GB | ~15 GB | Slow, 25 steps | Excellent |
flux-dev:q6 | Q6 | ~12 GB | ~20 GB | Slow, 25 steps | Best FLUX quality/size trade |
flux-dev:bf16 | BF16 | ~26 GB | ~36 GB | Slow, 25 steps | Best FLUX quality |
flux2-klein:q4 | Q4 | ~5 GB | ~11 GB | Fast, 4 steps | Good for very small GPUs |
flux2-klein:q8 | Q8 | ~6 GB | ~13 GB | Fast, 4 steps | Good |
z-image-turbo:q8 | Q8 | ~9 GB | ~13 GB | Fast, 9 steps | Excellent |
sdxl-turbo:fp16 | FP16 | ~8 GB | ~11 GB | Very fast, 4 steps | Good |
sd15:fp16 | FP16 | ~6 GB | ~6 GB | Medium, 25 steps | Good, broad ecosystem |
sd3.5-large:q8 | Q8 | ~12 GB | ~22 GB | Medium, 28 steps | Excellent |
qwen-image:q4 | Q4 | ~14 GB | ~22 GB | Slow, 50 steps | Good, validated at 1024 |
qwen-image-2512:q4 | Q4 | ~14 GB | ~22 GB | Slow, 50 steps | Good, validated at 1328 |
qwen-image:q8 | Q8 | ~22 GB | ~24+ GB | Slow, 50 steps | Best GGUF, validated at 768 |
Sequential vs Eager
In sequential mode (the default), mold loads each component (encoder → transformer → VAE) one at a time, freeing GPU memory between phases. This reduces peak VRAM by 30-50% but adds 10-20% to generation time.
Use --eager to keep all components loaded simultaneously for faster inference on high-VRAM cards. FLUX.1, FLUX.2, Z-Image, Qwen-Image, SD 3.5, LTX-2, and Wan also support --offload for block-level CPU↔GPU streaming (~24 GB down to ~2-4 GB peak, 3-5x slower).
Model Management
mold pull flux2-klein:q8 # Download a model
mold list # See what you have
mold info # Installation overview
mold info flux-dev:q4 # Model details + disk usage
mold rm dreamshaper-v8 # Remove a model
mold default flux-dev:q4 # Set default modelName Resolution
Bare names auto-resolve by trying :q8 → :fp16 → :bf16 → :fp8:
mold run flux2-klein "a cat" # resolves to flux2-klein:q8
mold run sdxl-base "a cat" # resolves to sdxl-base:fp16HuggingFace Auth
Some model repos (marked [gated]) require a HuggingFace access token. You may need to accept the model's license on its HuggingFace page before downloading.
Option 1: Environment variable (simplest):
export HF_TOKEN=hf_...
mold pull flux-dev:q4Option 2: HuggingFace CLI (persists the token):
# Install the HF CLI
curl -LsSf https://hf.co/cli/install.sh | bash
# Log in (saves token to ~/.cache/huggingface/)
hf auth loginOnce logged in, mold pull picks up the stored token automatically; no HF_TOKEN export needed.
See the HuggingFace CLI docs for more options.
All Families
| Family | Native Resolution | Architecture |
|---|---|---|
| FLUX.2 | 1024x1024 | Klein Qwen3 or Dev Mistral3 transformer family |
| FLUX.1 | 1024x1024 | Flow-matching transformer |
| SDXL | 1024x1024 | Dual-CLIP, UNet |
| SD 1.5 | 512x512 | CLIP-L, UNet |
| SD 3.5 | 1024x1024 | Triple encoder, MMDiT |
| Z-Image | 1024x1024 | Qwen3 encoder, 3D RoPE |
| Wuerstchen | 1024x1024 | 3-stage cascade, 42x compress |
| Qwen-Image | 1328x1328 | Qwen2.5-VL, flow-matching, CFG |
| Qwen-Image-Edit | Derived from first edit image | Qwen2.5-VL multimodal edit, flow-matching, CFG |
| LTX Video | 768x512 / 1216x704 | T5/Gemma, video transformers, causal VAEs |
| MiniMax H3 | 1344x768 | Qwen3-VL, joint audio-video DiT, dual VAEs |
| Wan Video | 832x480 / 1280x704 | UMT5-XXL, flow DiT, causal 3D VAE, A14B MoE |
| Hunyuan3D | n/a — output is a mesh | DINOv2-giant, flow DiT, vecset shape VAE |
| Upscalers | 2x / 4x source size | Real-ESRGAN super-resolution |
Each family page lists its actual shape contract. Bucketed families may warn when a request misses their recommended dimensions. MiniMax H3 instead accepts any 32-aligned canvas inside its documented continuous area and aspect bounds.
Maintainers should use the model resolution and aspect-ratio matrix for exact reduced ratios, checkpoint and pipeline exceptions, Mold admission bounds, profile hashes, and pinned upstream provenance.
Backend qualification
All image families and LTX Video run on CUDA, Apple Metal, and CPU. LTX-2 / LTX-2.3 is performance-qualified on CUDA and Apple Metal (measured on the 19B and 22B distilled FP8 tiers; Metal is slower than a comparable CUDA card). LTX-2.5's compact distilled INT8 ConvRot split pack is qualified on Apple Metal; Q3_K_M, Q4_K_M, and Q6_K GGUF are also Metal-qualified, while its BF16 route remains operator-deferred. CUDA has a separate completed qualification campaign. The LTX-2 family CPU path stays correctness-oriented and can be extremely slow. Wan is performance-qualified on CUDA; its CPU and Apple Metal paths are correctness-oriented (fp8-scaled Wan checkpoints stay CUDA-only; Metal has no fp8 widening kernel). MiniMax H3 compact checkpoints are downloadable everywhere, and both reviewed routes (FL2VA's boundary frame and Ref2VA's ordered image/video/audio references) execute on supported SM89 CUDA builds; the CPU path is unsupported, and the Apple Metal route is admitted and shipped but correctness-only and not yet hardware-qualified.