AuraFlow
The largest fully open-source flow-based text-to-image model, MMDiT under Apache 2.0.
About
AuraFlow is fal's open-source text-to-image model series, built by Simo Ryu and collaborators and released starting mid-2024 with v0.1. It is a multimodal diffusion transformer (MMDiT, in the Stable Diffusion 3 family of architectures) distinguished by shipping under Apache 2.0, at launch, the largest flow-based generation model with a fully permissive license, free to download, use, and monetize. The v0.2 checkpoint added more training compute; v0.3 fine-tuned on more aesthetic data and extended supported resolutions to 1536px on each side with varied aspect ratios. The HF model cards report state-of-the-art results on GenEval for the open-source class.
The honest position on AuraFlow is that it was overtaken quickly: contemporary community assessments rank it as outclassed on launch-era quality by the models that followed, and it is heavy to run on consumer hardware. But the Apache 2.0 license kept it relevant as a base for commercial fine-tunes where Stability-licensed models cannot go. fal hosts it on its API alongside the Hugging Face weights and diffusers support. If you want a flow-based DiT you can actually ship, AuraFlow is still the permissive-license option; if you want the best open weights of 2026, look elsewhere.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
State-of-the-art on GenEval (author-reported on HF model cards, v0.2/v0.3)
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 6/10
Scored 6/10 (metrics tier, metrics-v1). Evidence: fal/AuraFlow-v0.3 and v0.2 HF model cards (Apache 2.0, GenEval SOTA claim, 1536px support); diffusers AuraFlow pipeline docs; geeky-gadgets overview. Strengths: genuinely permissive Apache 2.0 license on a 6B-class flow model, fal-hosted API, diffusers support. Docked for: (1) GenEval numbers are author-reported with no independent verification found; (2) community consensus is that later open models outclassed it on quality and it is expensive to run locally. License: Apache-2.0.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 5.0/10
- Maintenance · recency, releases, issue load
- 5.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 6.0/10
- · fal; v0.1 mid-2024, v0.2/v0.3 later; MMDiT flow-based
- · GenEval SOTA is author-reported; community consensus is later open models outclassed it
- · Heavy on consumer hardware; fal API available