Skip to content
Bud's Wiki

fal open-sources FLUX.2 [dev] Turbo — distilled FLUX.2 with sub-second generation

TL;DR

Section titled “TL;DR”

fal announces the open-source release of FLUX.2 [dev] Turbo, an in-house distilled variant of Black Forest Labs’ FLUX.2 [dev]. The tweet claims #1 ELO among open-source image models on the Artificial Analysis arena, sub-second generation, and a “custom variant of DMD2 distillation for max quality.” No technical report or weights link is attached to the tweet itself — this page is a snapshot of the announcement.

Key claims

Section titled “Key claims”
  • FLUX.2 [dev] Turbo is fal’s in-house distilled version of the FLUX.2 [dev] base model [tweet].
  • The distillation recipe is described as a custom variant of DMD2 (Distribution Matching Distillation 2), tuned for quality rather than minimum-NFE [tweet].
  • The released model is positioned as the #1 ELO open-source image model on the Artificial Analysis text-to-image arena at time of posting [tweet].
  • Inference latency is sub-second per image [tweet].

Method

Section titled “Method”

Not disclosed in the tweet. The only architectural claim is “custom variant of DMD2 distillation” applied to FLUX.2 [dev] as teacher. No NFE count, no LoRA / full-finetune disclosure, no schedule details, no comparison numbers against other FLUX.2 distillations (e.g. FLUX.2 [klein]).

Results

Section titled “Results”

The tweet provides a single qualitative claim (#1 ELO open-source on Artificial Analysis arena) and a latency claim (sub-second). No quantitative comparison table, no per-step ablation, no diversity-vs-quality numbers. A single sample image is attached to the tweet.

Why it’s interesting

Section titled “Why it’s interesting”

The wiki’s Diffusion Distillation page tracks a cluster of 2025–2026 work showing that DMD-family distillation is now the production path to 1–4 NFE on FLUX-class models. FLUX.2 [dev] Turbo is a production datapoint for that thesis — fal applies a DMD2 variant, not the more recent ArcFlow / DP-DMD / TDM / TDM-R1 recipes the wiki has filed, suggesting either (a) the newer recipes haven’t fully diffused to production yet, or (b) DMD2 with quality-tuned hyperparameters is still the practical sweet spot for FLUX-class teachers. The “DMD2-for-quality” framing also contrasts with the Diversity-Preserved Distribution Matching Distillation for Fast Visual Synthesis (DP-DMD) thesis that vanilla DMD’s reverse-KL is mode-seeking and costs diversity — fal’s tweet does not address whether the Turbo variant preserves teacher diversity.

This is also the second open-weight FLUX.2 distillation the wiki has seen, after BFL’s own FLUX.2 [klein]: Towards Interactive Visual Intelligence (FLUX.2 [klein]) and the KV-cache-optimized FLUX.2 [klein] 9B-KV: KV-cache optimized variant for accelerated multi-reference editing variant — making FLUX.2 the most-distilled open image base of late 2025.

See also

Section titled “See also”