← sources

Flux2 Klein: cut the Qwen3 text encoder’s VRAM from 5.5 to 3.2 GB with identical output (new node)

source: reddit (reddit:StableDiffusion)published 2026-10-27 18:02ingested 2026-10-28 02:37

retained text

Flux2 Klein only reads three hidden layers from its Qwen3 text encoder (9, 18 and 27), but the standard loaders build and run the full 36-layer model, plus the final norm and the LM head. On top of that, ComfyUI-GGUF dequantizes the whole token embedding table to fp16 when it loads the GGUF.

I made a small loader that only builds what Klein actually uses:

* **27 layers instead of 36**, no final norm, no LM head
* **GGUF token embeddings stay quantized**: only the rows of your prompt are dequantized
* **Same conditioning, bit for bit** (compared against CLIPLoader (GGUF), max abs diff 0.0)

**Numbers (Qwen3-8B Q4\_K\_M, Klein 9B, RTX 4090, 512-token prompt):**

|\-|CLIPLoader (GGUF)|Klein CLIP Loader (Pruned)|
|:-|:-|:-|
|TE weights on GPU|5.51 GB|**3.20 GB**|
|Peak VRAM during encode|5.97 GB|**3.62 GB**|

That's \~2.3 GB back for the DiT, a LoRA stack, or an LLM running next to ComfyUI. On an 8 GB card it can be the difference between offloading or not.

**Install:**

    cd ComfyUI/custom_nodes
    git clone https://github.com/dchatel/comfyui_davcha_flux2klein

Then swap your text encoder loader for **Klein CLIP Loader (Pruned)** (category `davcha/flux2klein`). GGUF files need ComfyUI-GGUF installed. No other dependencies.

Tested with Qwen3-8B Q4\_K\_M (Klein 9B). The Qwen3-4B (Klein 4B) and safetensors paths use the same settings but I haven't verified them myself yet. Reports welcome.

Repo: [https://github.com/dchatel/comfyui\_davcha\_flux2klein](https://github.com/dchatel/comfyui_davcha_flux2klein)

Side note: while profiling this I also noticed ComfyUI picks which model to unload by reference count and size, not by how long the model takes to reload. I opened a PR that ranks unload candidates by measured reload time instead (13.6 s → 6.7 s per run in a Klein + local LLM workflow): [https://github.com/Comfy-Org/ComfyUI/pull/17363](https://github.com/Comfy-Org/ComfyUI/pull/17363) (not merged, feedback welcome).

document fetched 2026-10-28 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1xgzujz/flux2_klein_cut_the_qwen3_text_encoders_vram_from