Quantised vs Full-Precision Flux on 12GB to 24GB GPUs

Running Flux locally comes down to one stubborn fact: the full model is large, and your VRAM is finite. Quantised versus full-precision Flux is the lever that decides whether a 12GB card runs it at all. Flux.1 Dev at full FP16 wants roughly 24GB of VRAM, but FP8 or Q8 GGUF quantisation squeezes it to around 12GB, and Q4 drops it near 6 to 8GB. What you give up in exchange is image fidelity, and how much you give up depends heavily on which level you choose.

Quick Answer

Flux.1 Dev needs around 24GB at FP16. FP8 or Q8 GGUF roughly halves that to about 12GB with quality almost nobody can tell apart from full precision. Q4 GGUF brings it down near 6 to 8GB but softens fine detail and weakens text rendering. On a 12GB card, FP8 or Q8 is the sweet spot; on 24GB, run full FP16 and stop worrying.

What each precision level costs you

FP16 full precision

This is the reference: the model exactly as released, no compression. It needs roughly 24GB of VRAM in practice once the text encoder and working memory are accounted for, which puts it firmly in 24GB-class territory. The reward is maximum fidelity, the cleanest fine detail and the most reliable text rendering. The cost is that anything below a 24GB card cannot hold it without spilling over and crawling.

FP8 quantisation

FP8 halves the memory footprint to around 12GB while keeping image quality within a whisker of FP16. In blind comparisons most people simply cannot pick FP8 apart from full precision, which is why it is the default recommendation for 16GB cards and the entry ticket for 12GB ones. It is a single, easy-to-load file and the best balance of quality and footprint for most local setups.

GGUF Q8 down to Q4

GGUF quantisation goes further and more granularly. Q8 is near-lossless and behaves much like FP8 on memory. Step down through the levels and the savings grow while quality slowly erodes, until Q4 fits the model in roughly 6 to 8GB. At Q4 the trade becomes visible: fine details soften and text rendering gets less accurate. For quick drafts, composition tests and concept work on a small card, that is an acceptable price. For final, detailed output you would not stop there if you can avoid it.

Matching the model to your card

The decision is really driven by your GPU. On a 24GB card, run FP16 and accept no compromise on quality or resolution. On a 16GB card, FP8 (or Q8 GGUF) is the comfortable sweet spot, near-full quality at half the memory. On a 12GB card, FP8 still fits and is the recommended target; you can drop to a lower GGUF level if you need more headroom for higher resolutions or extra nodes. Below that, on 8 to 10GB, Q4-class GGUF is what makes Flux run at all, with the detail trade-off baked in.

ComfyUI is the practical home for all of this, since it supports every format, FP16, FP8 and the full GGUF range, and lets you swap precision without re-architecting your workflow. If you are choosing a card with Flux specifically in mind, the graphics card range at Evetech spans the 12GB to 24GB band where these decisions live.

GGUF vs FP8: Which Format to Actually Use

The two formats are not always interchangeable. FP8 is a single fixed-point format: the model is stored as one file and every weight is quantised to 8-bit floating point. GGUF is a container format that supports a range of integer quantisation levels from Q2 through Q8, each with different memory footprints and quality trade-offs. For most local setups the practical difference is small: GGUF Q8 and FP8 produce results that are effectively indistinguishable in typical prompt comparisons. Where GGUF earns its place is on smaller cards, because it can compress further than FP8 allows, and because community model releases often ship only as GGUF. FP8 earns its place on cards with 16GB or more, where the single-file simplicity and consistent quality land without the memory pressure that forces you down through the GGUF levels.

Flux.2 Changes the Maths at the Top

The original Flux.1 Dev is a 12-billion-parameter model. Flux.2, released by Black Forest Labs in late 2025, doubles that to 32 billion parameters. At full precision it needs roughly 90GB, and even FP8 requires around 32GB. For 12GB card owners Flux.2 is GGUF-only territory at very aggressive compression, and the quality trade-off at those levels is more visible than it is with Flux.1. Knowing this matters because the practical choice for 12GB hardware is still Flux.1 Dev at FP8 or Q8 GGUF; that remains the recommended target and is not made obsolete by the arrival of Flux.2.

The practical rule

Use the highest precision your VRAM holds comfortably, and no lower. FP8 is invisible enough that paying the storage for FP16 only makes sense once you genuinely have 24GB. Q4 GGUF is a tool for fitting a model onto a small card, not a quality choice, so reach for it when capacity forces your hand rather than by default. Q5_K_S is considered a better stopping point than Q4 if your card has the headroom, since it is nearly as compact but meaningfully sharper on fine detail. If you are buying hardware for this seriously, the best selling GPUs at Evetech are a useful shortlist of cards that land in the right VRAM tiers.

Frequently Asked Questions

How much VRAM does Flux.1 Dev need at full precision?

Around 24GB once the text encoder and working memory are included, which is why full FP16 is comfortable only on 24GB-class cards. Below that you need a quantised version.

Is FP8 Flux noticeably worse than FP16?

For most prompts, no. FP8 keeps quality within a whisker of full precision and is very hard to distinguish in blind comparisons, which is why it is the standard recommendation for 12GB and 16GB cards.

What does Q4 GGUF actually sacrifice?

Q4 fits Flux into roughly 6 to 8GB but softens fine detail and makes text rendering less accurate. It is best for drafts and composition tests on small cards rather than final detailed images.

Which Flux format should I run on a 12GB GPU?

FP8 or Q8 GGUF is the target, near-full quality at about 12GB. Drop to a lower GGUF level only if you need extra headroom for higher resolutions or heavier workflows.

Does ComfyUI support all these formats?

Yes. ComfyUI handles FP16, FP8 and the full GGUF range, so you can switch precision to match your card without rebuilding your workflow.