Home / Hardware compatibility
The right plan for your machine.
Know what can run before you download. UnstableField plans around actual RAM and VRAM—not a one-size-fits-all GPU recommendation.
From 4–6 GB NVIDIA
Small 512px profiles are estimated. The 8 GB profile supports quantized 1024px images.
Start with desktop NVIDIA
12–16 GB VRAM and 32 GB+ RAM. Model choice is gated by RAM; not every video model fits.
GPU or CPU
Qwen3 8B on listed GPU profiles. CPU: Qwen3 4B at 8 GB RAM or 8B at 10 GB RAM.
These are task and memory profiles, not speed benchmarks. “Calibrated,” “measured under cap” and “estimated” are different kinds of evidence.
NVIDIA profiles
Start with VRAM. Check RAM, too.
Image quality, resolution and video eligibility depend on the model and the memory available to its plan.
NVIDIA desktop · 12–16 GB
12–16 GB VRAM / 32 GB+ RAM
IMAGE
Z-Image Turbo: 1024px, bf16_offload with 35 GB+ RAM; otherwise gguf_q8. FLUX.2 klein: 768px on 12 GB, 1024px on 16 GB.
VIDEO
Wan2.2 5B: bf16_offload with 36 GB+ RAM; otherwise gguf_q8. A14B Lightning needs 52 GB+ RAM.
CHAT
Qwen3 8B
Image + video. Video model eligibility still depends on RAM.
NVIDIA · 8 GB
8 GB VRAM / 16 GB RAM
IMAGE
Z-Image Turbo and FLUX.2 klein: 1024px using gguf_q4.
VIDEO
No video.
CHAT
Qwen3 8B
Measured on a capped desktop, not a real laptop. A real laptop will be slower.
NVIDIA · 4–6 GB
4 or 6 GB VRAM / 16 GB RAM
IMAGE
Small profile: FLUX.2 klein at 512px with reduced gguf_q4_small. Z-Image Turbo is also eligible with the full 16384 MB RAM.
VIDEO
No video.
CHAT
Qwen3 8B
Both 4 GB and 6 GB profiles use the same Small profile. Not yet run on these devices.
Other platforms
Useful starting points. Honest limits.
CPU chat and Apple, ROCm and XPU profiles are estimated. Experimental video does not mean measured or calibrated support.
CPU only
8 GB or 10 GB RAM
IMAGE
No image generation.
VIDEO
No video.
CHAT
Qwen3 4B with 8 GB RAM; Qwen3 8B with 10 GB RAM.
Chat only. Chat GGUF eligibility is gated by RAM, not VRAM.
Apple Silicon · MPS
16 GB memory profile
IMAGE
Z-Image Turbo and FLUX.2 klein: gguf_q4 at 1024px.
VIDEO
No MPS video support listed.
CHAT
Qwen3 8B
Image support is estimated, not measured on this platform.
AMD · ROCm
16 GB profile
IMAGE
Z-Image Turbo: gguf_q8. FLUX.2 klein: bf16.
VIDEO
Experimental with gguf_q8.
CHAT
Qwen3 8B
Other-platform estimates are starting points, not measured guarantees.
Intel Arc · XPU
16 GB profile
IMAGE
Z-Image Turbo: int8. FLUX.2 klein: bf16.
VIDEO
Experimental with int8.
CHAT
Qwen3 8B
Other-platform estimates are starting points, not measured guarantees.
Hardware-aware, not hardware magic
The planner handles the technical choices.
Actual VRAM and RAM drive weight precision, offloading and attention. Less manual tuning—not a promise that every model fits.
Weight precision
Quantized plans such as gguf_q4 reduce memory pressure. The Small profile also reduces resolution to 512px. Eligibility is model-specific.
Offloading
Moving weights out of VRAM needs system RAM. That is why bf16_offload has explicit RAM thresholds and larger video models need more.
Attention and GPU sharing
The planner chooses attention for the available memory. Chat uses its own text runtime and shares the card with generation jobs.
Read the evidence
A profile is only as strong as its basis.
The measurements come from RTX 4070 Ti work. Distinguish runs on the named card from capped tests and formula-based estimates.
Run on the named card. Thresholds use the measured device peak plus 2048 MB.
RTX 4070 Ti capped to 8 GB VRAM and 16 GB RAM. VRAM thresholds use allocator reserved peak + 2048 MB; RAM uses anonymous RSS + 2048 MB. This is not a real laptop test.
Not run on the target device. Values come from the planner formula or a starting VRAM value. Treat these profiles as starting points, not verified performance.
Check the model, not just the machine.
Model licenses vary. Open weights do not imply unrestricted commercial use. Review each model’s license and requirements. HunyuanVideo 1.5 and LTX-2.5 runners are unmeasured—not calibrated recommendations.
Models, tiers and licenses ↗
View on GitHub