aether-models/flux2-klein-4b

FLUX.2 [klein] 4B

A Core AI bundle of black-forest-labs/FLUX.2-klein-4B for the Aether SDK (iOS and macOS 27+).

• Source: black-forest-labs/FLUX.2-klein-4B at revision e7b7dc27f91deacad38e78976d1f2b499d76a294, licence Apache-2.0. The licence is included as LICENSE.
• Changes from the source: converted from PyTorch to Core AI (.aimodel) by Aether forge (recipe flux2-klein-4b@3). Weights are int8-linear-perchannel (8-bit weights). Three assets per variant, run one after another: the text encoder (these weights, cut after decoder layer 27), the denoiser (these weights, float32 residual stream) and the image decoder (float16 weights, float32 compute, never quantized). The tokenizer files are the source's own.
• Output: images from text prompts (imageOutput), 4 steps by default, no guidance.

Variants

Variant	Platform	Arch	Compute	Compiled	Assets	Download
macos-any-gpu	macos	any	gpu	no (specialized on first load)	flux2_klein_4b-denoiser0.aimodel 1.9 GB, flux2_klein_4b-denoiser1.aimodel 1.28 GB, flux2_klein_4b-denoiser2.aimodel 1.44 GB, flux2_klein_4b-text.aimodel 3.12 GB, flux2_klein_4b-vae.aimodel 101.5 MB	7.87 GB
ios-any-gpu	ios	any	gpu	no (specialized on first load)	flux2_klein_4b-denoiser0.aimodel 1.9 GB, flux2_klein_4b-denoiser1.aimodel 1.28 GB, flux2_klein_4b-denoiser2.aimodel 1.44 GB, flux2_klein_4b-text.aimodel 3.12 GB, flux2_klein_4b-vae.aimodel 101.5 MB	7.87 GB

Image sizes:

• macos-any-gpu: 1024×1024 (default), 512×512, 768×768, 1024×768, 768×1024, 1344×768, 768×1344
• ios-any-gpu: 512×512 (default), 576×448, 448×576, 768×768, 1024×576, 576×1024, 1024×1024

ios-any-gpu needs the com.apple.developer.kernel.increased-memory-limit entitlement: its T2 run peaked at 4590 MB on iPhone18,2 (record e10b51c0). Its stages load one at a time.

Verification

Every row is a record in verification/ about exactly these bytes (matched by bundle digest). Reference rows are strict T2 passes of the unquantized export on the same fixture, in verification/reference/.

Variant	Tier	Result	Detail	Device	OS build	Compute	Record
ios-any-gpu	T2	pass	8/8 cases; profile quantized-8bit; image path SSIM ≥ 0.94; fixture 523d996477835636	iPhone18,2	24A437	target	e10b51c0
macos-any-gpu	T2	pass	8/8 cases; profile quantized-8bit; image path SSIM ≥ 0.94; fixture 523d996477835636	Mac17,6	26A428	target	2dbe813e
macos-any-gpu	T3	pass	imagegen-shapes-v1; 100.0% vs reference 100.0%; 100 items; judged by qwen3-vl-2b	Mac17,6	26A428	target	10a8baac
unquantized reference (not published)	T2	pass	8/8 cases; profile strict; image path SSIM ≥ 0.98; fixture 523d996477835636	Mac17,6	26A428	target	70d5bf8d

The quantized profile also requires: T2 strict on the unquantized reference export (met by the reference row).

Use

aether generate flux2-klein-4b --prompt "A lighthouse at dusk, watercolor" -o lighthouse.png

import Aether

let aether = try Aether()
let generator = try await aether.imageGenerator("flux2-klein-4b")
let image = try await generator.image(for: ImageRequest(prompt: "A lighthouse at dusk, watercolor"))
try image.pngData().write(to: URL(fileURLWithPath: "lighthouse.png"))