Models
mold supports a broad range of model families spanning different architectures, quality levels, and VRAM requirements — including both image and video generation.
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 |
| Low VRAM (<8 GB) | flux2-klein:q4 | 2.6 GB, 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, 1024x1024 |
| Video | ltx-video-0.9.6-distilled:bf16 | Text-to-video, 30fps, APNG/MP4, best-supported default |
| Audio + video | ltx-2-19b-distilled:fp8 | Joint audio-video, MP4-first, advanced conditioning |
VRAM Guide
| Model | Variant | Approx. VRAM | Speed | Quality |
|---|---|---|---|---|
flux-schnell:q8 | Q8 | ~12 GB | Fast, 4 steps | Good |
flux-schnell:q6 | Q6 | ~14 GB | Fast, 4 steps | Better than Q8 |
flux-dev:q4 | Q4 | ~8 GB | Slow, 25 steps | Excellent |
flux-dev:q6 | Q6 | ~10 GB | Slow, 25 steps | Best FLUX quality/size trade |
flux-dev:bf16 | BF16 | ~24 GB | Slow, 25 steps | Best FLUX quality |
flux2-klein:q4 | Q4 | ~4 GB | Fast, 4 steps | Good for very small GPUs |
z-image-turbo:q8 | Q8 | ~10 GB | Fast, 9 steps | Excellent |
sdxl-turbo:fp16 | FP16 | ~10 GB | Very fast, 4 steps | Good |
sd15:fp16 | FP16 | ~6 GB | Medium, 25 steps | Good, broad ecosystem |
qwen-image:q4 | Q4 | ~14 GB | Slow, 50 steps | Good, stable at 1024x1024 |
qwen-image-2512:q4 | Q4 | ~14 GB | Slow, 50 steps | Good, stable at 1024x1024 |
qwen-image:q8 | Q8 | ~22 GB | Slow, 50 steps | Best GGUF, validated at 768 |
qwen-image-flash:q4 | Q4 | ~13 GB | Fast, 4 steps | DMD2 distill; softer on fine detail |
qwen-image-distill:q4 | Q4 | ~14 GB | Medium, 15 steps | Distill-Full merge, closer to base |
ltx-video-0.9.6-distilled:bf16 | BF16 | ~10 GB | Fast, 8 steps | Video, low-VRAM default |
ltx-video-0.9.8-2b-distilled:bf16 | BF16 | ~10-12 GB | Fast, 7+3 steps | Newer video checkpoint, multiscale refine |
ltx-2-19b-distilled:fp8 | FP8 | ~24 GB | Slow, 8 steps | Joint audio-video, recommended LTX-2 |
ltx-2.3-22b-distilled:fp8 | FP8 | ~24 GB | Slow, 8 steps | Larger joint audio-video path |
ltx-2.3-22b-dev:bf16 | BF16 | 48 GB+ | Very slow, streamed below 48 GB | Full-quality trainable reference path |
ltx-2.3-22b-distilled:bf16 | BF16 | 48 GB+ | Slow, streamed below 48 GB | Full-precision eight-step path |
VRAM 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 |
sd35-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 1024 |
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 also supports --offload for block-level CPU↔GPU streaming (~4-5 GB peak, 2-4x 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-2 | 1216x704 | Gemma 3, joint audio-video transformer |
| LTX Video | 768x512 | T5-XXL, DiT, 3D causal VAE |
| MiniMax H3 | 1344x768 | Qwen3-VL, joint audio-video DiT, dual VAEs |
| Wan Video | 832x480 / 1280x704 | UMT5-XXL, flow DiT, causal 3D VAE, A14B MoE |
Each family page lists recommended dimensions for non-square aspect ratios. Using non-recommended dimensions will trigger a warning.
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); its 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, but only the reviewed FL2VA route can execute, on supported SM89 CUDA builds; Ref2VA execution, Metal, and CPU are unsupported.