mannix/ flux2-klein

393 Downloads Updated 1 week ago

4B diffusion model · Qwen3-4B text encoder · FLUX.2 VAE · 1024×1024 in four steps

4b

ollama run mannix/flux2-klein:4b

curl http://localhost:11434/api/chat \
  -d '{
    "model": "mannix/flux2-klein:4b",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from ollama import chat

response = chat(
    model='mannix/flux2-klein:4b',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'mannix/flux2-klein:4b',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Models

Name

1 model

Size / Usage

Context

Input

flux2-klein:4b

7.0GB · - context window · Text · 1 week ago

flux2-klein:4b

7.0GB

-

Text

Readme

FLUX.2 klein 4B — image generation and reference edits, for xOllama

4B diffusion model (Q8_0 GGUF) · Qwen3-4B text encoder · FLUX.2 VAE · 1024×1024 in four steps · OpenAI /v1/images routes

Runs on xOllama, not on stock ollama. Stock ollama pulls this model but cannot show or run it: it carries only media engines. xOllama serves it through an opencoti engine with media support (dev build 2610022033001/b105 or later, set with XOLLAMA_ENGINE_PATH). The engine pinned by the current xOllama release has no media engines and refuses the model at load, naming what it lacks.

FLUX.2 klein generates and edits images in four steps, in the least memory of the image models here. Edits are reference edits: the input images steer the result instead of being noised over (the engine picks this for FLUX.2 by itself).

Use

xollama pull mannix/flux2-klein:4b
curl http://localhost:22434/v1/images/generations \
  -d '{"model": "mannix/flux2-klein:4b", "prompt": "a lighthouse at dusk, oil painting"}'

An edit is multipart, with one or more image[]:

curl http://localhost:22434/v1/images/edits \
  -F model=mannix/flux2-klein:4b -F prompt="make it winter" -F "image[]=@photo.png"

Any OpenAI client works: the reply is b64_json, or a data: URL with "response_format": "url". Besides OpenAI’s fields the engine takes steps, cfg_scale, sampler, scheduler, seed and negative_prompt.

What the template fills in

A client that sends only a prompt gets the template’s settings: 1024×1024, 4 steps, cfg 1, euler, png. An edit without a size comes back at its first image’s size (scaled to at most 1024×1024’s pixel count, in multiples of 16).

Measured

	RTX 3090
1024×1024 generation	16 s (opencoti b97); 7.6 s warm (b105)
reference edit, 1024×1024	41.2 s (b105)

Tags

Tag	Weights	Size
`4b`	`flux-2-klein-4b-Q8_0.gguf` + Qwen3-4B Q4_0 + FLUX.2 VAE	7.0 GB

Source

GGUF: leejet/FLUX.2-klein-4B-GGUF · VAE: Comfy-Org/vae-text-encorder-for-flux-klein-4b · text encoder: unsloth/Qwen3-4B-GGUF · License: Apache-2.0. Docs: xOllama media.

# FLUX.2 klein 4B — image generation and reference edits, for xOllama

**4B diffusion model (Q8_0 GGUF) · Qwen3-4B text encoder · FLUX.2 VAE · 1024×1024 in four steps · OpenAI `/v1/images` routes**

> **Runs on [xOllama](https://github.com/mann1x/xollama), not on stock ollama.** Stock ollama pulls this model but cannot show or run it: it carries only media engines. xOllama serves it through an [opencoti](https://huggingface.co/ManniX-ITA/opencoti-llamafile) engine with media support (dev build `2610022033001`/b105 or later, set with `XOLLAMA_ENGINE_PATH`). The engine pinned by the current xOllama release has no media engines and refuses the model at load, naming what it lacks.

FLUX.2 klein generates and edits images in four steps, in the least memory of the image models here. Edits are reference edits: the input images steer the result instead of being noised over (the engine picks this for FLUX.2 by itself).

## Use

```shell
xollama pull mannix/flux2-klein:4b
curl http://localhost:22434/v1/images/generations \
  -d '{"model": "mannix/flux2-klein:4b", "prompt": "a lighthouse at dusk, oil painting"}'
```

An edit is multipart, with one or more `image[]`:

```shell
curl http://localhost:22434/v1/images/edits \
  -F model=mannix/flux2-klein:4b -F prompt="make it winter" -F "image[]=@photo.png"
```

Any OpenAI client works: the reply is `b64_json`, or a `data:` URL with `"response_format": "url"`. Besides OpenAI's fields the engine takes `steps`, `cfg_scale`, `sampler`, `scheduler`, `seed` and `negative_prompt`.

## What the template fills in

A client that sends only a prompt gets the template's settings: **1024×1024, 4 steps, cfg 1, euler, png**. An edit without a size comes back at its first image's size (scaled to at most 1024×1024's pixel count, in multiples of 16).

## Measured

| | RTX 3090 |
|---|---|
| 1024×1024 generation | 16 s (opencoti b97); 7.6 s warm (b105) |
| reference edit, 1024×1024 | 41.2 s (b105) |

## Tags

| Tag | Weights | Size |
|---|---|---|
| `4b` | `flux-2-klein-4b-Q8_0.gguf` + Qwen3-4B Q4_0 + FLUX.2 VAE | 7.0 GB |

## Source

GGUF: [leejet/FLUX.2-klein-4B-GGUF](https://huggingface.co/leejet/FLUX.2-klein-4B-GGUF) · VAE: [Comfy-Org/vae-text-encorder-for-flux-klein-4b](https://huggingface.co/Comfy-Org/vae-text-encorder-for-flux-klein-4b) · text encoder: [unsloth/Qwen3-4B-GGUF](https://huggingface.co/unsloth/Qwen3-4B-GGUF) · License: Apache-2.0.
Docs: [xOllama media](https://github.com/mann1x/xollama/blob/dev/docs/xollama/media.mdx).

Paste, drop or click to upload images (.png, .jpeg, .jpg, .svg, .gif)