FLUX.2-klein-4B-bf16

mlx-community/FLUX.2-klein-4B-bf16

MLX (bf16) snapshot of black-forest-labs/FLUX.2-klein-4B (Apache-2.0) for Apple Silicon — a compact rectified-flow MMDiT (5 double + 20 single blocks)

• Qwen3-4B 3-layer-tap conditioner + FLUX.2 VAE. Standard diffusers tree (transformer/ text_encoder/ vae/ tokenizer/), loaded by the Swift/MLX port; int8/int4 are produced at load time (int4 DiT ~2.35 GB → a q4 pipeline fits a 16 GB Mac).

4B only — the 9B klein variants are FLUX Non-Commercial; this Apache-2.0 4B is the wrapped tier.

Parity (Swift port vs diffusers goldens, fp32/CPU)

• MMDiT: cosine ≥0.9999995 at the 64×64 production grid (all blocks + norm_out + whole-forward)
• FLUX.2 VAE decode 105–130 dB · Qwen3 3-layer-tap encoder cosine 0.9999999
• GPU int4 1024²/4-step ~6 s.

Use (Swift / MLXEngine)

import MLXKlein
import MLXToolKit
let pkg = Klein4BT2IPackage(configuration: .init(quant: .int4, snapshotPath: "<this repo dir>"))
try await pkg.load()
let r = try await pkg.run(T2IRequest(prompt: "a red fox in a snowy forest at sunrise, photorealistic",
                                     width: 1024, height: 1024, seed: 42)) as! T2IResponse

Code: https://github.com/xocialize/flux2-klein-swift