FLUX.2 [klein] 4B for Geode

The image model of Geode's Images screen: a converted copy of FLUX.2 [klein] 4B by Black Forest Labs (a 4-step distilled rectified-flow transformer, 3.9 B parameters, with the Qwen3-4B text encoder), split into one file per stage so the phone loads the text encoder, runs it and frees it before the transformer loads, and the transformer before the VAE (only the largest stage is ever in memory). Geode's own engine runs it on the phone's processor, graphics chip or Snapdragon NPU; nothing leaves the phone.

Two folders; the app downloads one:

Folder Transformer (DiT) Download Quality against the original (fp32, seed 42, 512ยฒ)
fp16/ (default) fp16, 7.2 GiB 10.2 GB the same image (76 dB PSNR, SSIM 1.000)
int8/ (opt-in, lower quality) int8 per 32 inputs with a Hadamard-32 turn ("i8r"), 4.0 GiB 6.7 GB the same picture with fine texture differing (30-41 dB, SSIM 0.97-0.997)

Both share the same text encoder (Q4), VAE (fp16) and tokenizer files (byte for byte).

File What it is
tokenizer.bin Qwen3's byte-level BPE (the checkpoint's tokenizer/) in Geode's format
embed.bin the text encoder's token embeddings, bf16 as in the checkpoint (one row read per token)
te_q4.bin the text encoder's layers 0-26 in 4 bits (an E4M3 scale per 16 values); its norms fp32
dit_f16.bin / dit_i8r.bin the transformer: fp16, or its blocks' matrices in int8 (the small modulation / embedding / output layers stay fp16)
vae_f16.bin the VAE's decoder and the latents' batch-norm statistics, fp16 (norms and biases fp32)
vae_enc_f16.bin the VAE's encoder (+ quant_conv), fp16: reference images for editing (added 2026-11-18)
taef2_f16.bin TAEF2's tiny decoder (madebyollin/taef2, MIT), fp16: the live preview after each step (added 2026-11-18)
vae_small_f16.bin BFL's FLUX.2 small decoder (Apache-2.0), fp16: an opt-in faster decoder, downloaded only when turned on; its own pack in geode.json's vae_small (added 2026-11-18)
geode.json the pack's manifest: the configuration Geode reads, every file's size and SHA-256, every tensor
LICENSE the Apache License 2.0, as the original model ships it
LICENSE-TAEF2 TAEF2's MIT licence (Copyright (c) 2023 Ollin Boer Bohan)

Sources and changes (Apache-2.0, section 4)

  • Original: black-forest-labs/FLUX.2-klein-4B at revision e7b7dc27f91deacad38e78976d1f2b499d76a294, Copyright Black Forest Labs, licensed under the Apache License 2.0 (LICENSE copied here unchanged).
  • Changed by the Geode project (the files here are derived works, converted by Geode's tools/pack_image.py):
    • the transformer's bf16 weights stored as fp16 (fp16/; relative error ~1e-8), or its blocks' matrices and context_embedder quantized to int8 with one fp16 scale per 32 inputs after a normalized 32-point Walsh-Hadamard transform of each group (int8/);
    • the text encoder (Qwen3-4B) cut to the layers the pipeline reads (layers 0-26 of 36: FLUX.2 [klein] takes hidden states 9, 18 and 27; layers 27-35, the final norm and the LM head are left out) and quantized to 4 bits;
    • the VAE's decoder as fp16 rows, and (since 2026-11-18, for editing with up to 4 reference images) its encoder and quant_conv as fp16 in vae_enc_f16.bin;
    • BFL's FLUX.2 small decoder (Apache-2.0, Copyright Black Forest Labs): its decoder as fp16 rows in vae_small_f16.bin, unchanged otherwise;
    • TAEF2 by Ollin Boer Bohan (MIT, LICENSE-TAEF2): its decoder as fp16 rows in taef2_f16.bin;
    • the tokenizer re-encoded in Geode's binary format; no weight was trained or fine-tuned.
  • The pipeline (chat template, padding to 512 tokens, FlowMatchEuler scheduler with the empirical shift, torch's CPU noise for a seed) follows diffusers' Flux2KleinPipeline, so a seed gives the original pipeline's image up to the precision above.

Use

Black Forest Labs' model card lists the uses the model and its derivatives are not meant for: among them anything that violates applicable law, exploiting or harming minors, deceptive or fraudulent content, harassment, and non-consensual intimate imagery. Those apply to these files too. Nothing here changes the licence or adds terms of its own.

Speed

Measured per stage by Geode's engine and shown in the app (Images screen, the result's details); the targets are 512ยฒ in about 10 s on a Snapdragon 8 Elite (the NPU) and 2-3 minutes on a Snapdragon 720G (the processor).

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for Maximilian228/Geode-FLUX.2-klein-4B

Finetuned
(93)
this model