The 12B text-to-image transformer that reset the bar for open-weight image quality — non-commercial licensed, T5-powered prompt understanding, and a VRAM appetite that scales from 8GB quantized up to 24GB at full precision.
Last verified 2026-07 · Page created; VRAM tiers verified against Black Forest Labs’ own guidance and community GGUF (city96) quantization reports rather than an invented quant table.
FLUX.1 [dev] is Black Forest Labs’ 12B-parameter rectified-flow transformer, released August 2024, and it is widely regarded as the strongest open-weight text-to-image model for prompt adherence and photorealism — a large part of that comes from pairing it with a T5-XXL text encoder (up to 512 tokens) instead of the shorter CLIP-only prompting older models rely on. The license is non-commercial: personal and research use is fine, but commercial use requires a separate license from Black Forest Labs.
At full BF16/FP16 precision the transformer plus T5 text encoder needs roughly 24GB of VRAM to run without offloading — this is the "just works, full quality" tier, comfortable on a single RTX 3090/4090 or better. GGUF-quantized versions of the transformer (the same Q4/Q5/Q8 idea as an LLM, applied to the image model’s weights) bring that down substantially: Q8 fits comfortably in 12GB with minimal quality loss, and Q4 gets it running on 8GB cards, trading some fine detail for accessibility.
FP8 quantization (distinct from the GGUF ladder, common in ComfyUI workflows) is another well-supported middle ground, roughly halving the full-precision VRAM need while staying close to full quality.
Image models don’t quantize the same way LLMs do — practical VRAM guidance instead:
Unified-memory machines — Apple Silicon Mac Studio/MacBook Pro (Max/Ultra chips) and AMD Strix Halo mini-PCs — share system RAM and GPU memory in one pool, so a 64–128 GB configuration can hold models that would otherwise need multiple discrete GPUs. Throughput is bandwidth-bound like any GPU; it isn’t modeled in the table above yet.
Three ways to size a machine for FLUX.1 [dev], from the smallest GPU that fits to the no-compromise option with headroom for the next model up.
Owning the hardware is a one-time cost that keeps paying off the more you actually use it — every run after the first is free, the data never leaves your machine, and the GPU still does everything else a PC does (gaming, editing, the rest of your workload) between AI sessions. Renting cloud GPU time trades that upfront cost for pay-as-you-go pricing with no idle hardware to maintain, which wins for occasional, bursty, or much-larger-than-consumer-VRAM jobs. For a model that fits comfortably in the VRAM tiers above and gets used regularly, owning the GPU tends to be the cheaper path over time — for anything larger than a single card can hold, renting is often the more practical starting point.
Not at full precision (that needs roughly 24GB), but yes with a GGUF Q8 quantized transformer — that is the recommended tier for a 12GB card, keeping close to full-precision quality.
With aggressive GGUF quantization (Q4-class) and CPU offload for the text encoder, 8GB is workable, with a noticeable quality hit versus Q8 or full precision. Below 8GB is not a realistic target for this model.
Not under the standard [dev] license — it is non-commercial only. Black Forest Labs offers separate commercial licensing, and their FLUX.1 [pro] and [schnell] variants have different license terms if commercial use is the goal.