FLUX.2 [klein] 4B Edit -- prompt encoder (Qwen3-4B, CPU)

The text encoder of FLUX.2 [klein] 4B (Qwen3-4B, Apache-2.0) for on-device use, shared by all latentdivergence/flux2-klein-edit-npu-* builds.

text/qwen3_te.onnx (+ .data): the first 27 decoder layers of Qwen3-4B, every Linear as ONNX Runtime MatMulNBits (8-bit, block 32). Inputs inputs_embeds [1,64,2560] and attention_mask [1,64] (right padded); output prompt_embeds [1,64,7680] = hidden states after layers 9, 18 and 27, concatenated, exactly what diffusers' Flux2KleinPipeline._get_qwen3_prompt_embeds produces.
text/embed_q8.bin + embed_scale.bin: the token embedding table, int8 per row with float32 scales (row-major [151936, 2560]); look up the ids on the host.
text/qwen_tokenizer.bin: Qwen2 byte-level BPE (vocab, merges, added tokens) in a compact binary for the app.
text/qwen3_te.json: shapes and settings.

Prompt format: the Qwen3 chat template with one user turn, generation prompt, thinking disabled, padded to 64 tokens.

Licence: Apache-2.0 (see LICENSE.txt).

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. 🙋 Ask for provider support

Model tree for latentdivergence/flux2-klein-edit-text

Base model

Finetuned

Quantized

(348)

this model