<p>I wanted to see if Anima could run on <strong>Qwen3.5-0.8B</strong> instead of its stock Qwen3-0.6B text encoder, without retraining Anima itself. This is the result: Qwen3.5-0.8B-Base cut down to 20 layers, plus a small head (about 2.1M parameters) that translates its output into what Anima's LLM adapter expects.</p>

<p><strong>Think of it as an alternative, not an upgrade.</strong> On held-out prompts the adapter output lands within 2.2% of the 0.6B (cosine 0.988), and on 300 held-out images the denoising loss is the same (0.13346 vs 0.13347). In a blind test where I didn't know which side was which, I picked the 0.8B in 4 of 12 pairs, the 0.6B in 6, and 2 were ties. So it behaves like the stock encoder, with a newer model and a bigger tokenizer (248k vocab vs 151k) underneath.</p>

<h3>How to use</h3>
<ul>
<li><strong>Diffucore v0.2.1 or newer</strong>: put the file in <code>models/text-encoders/</code> and pick it in the <strong>TE</strong> dropdown. Older versions had a broken Qwen3.5 encoder, so update first.</li>
<li><strong>ComfyUI</strong>: the built-in loader can't read this file. Install the <a href="https://github.com/nawka12/comfyui-qwen35-anima-head" rel="ugc">comfyui-qwen35-anima-head</a> node, put the file in <code>models/text_encoders/</code>, and use <strong>Load Qwen3.5 Anima Text Encoder</strong> in place of the stock CLIP loader. Everything else in your Anima workflow stays the same.</li>
</ul>

<h3>How it was made</h3>
<p>Text-only distillation: the 0.6B is the teacher, both encoders go through Anima's frozen LLM adapter, and only the head learns to match the adapter output. The Qwen3.5 layers are untouched. 25,152 SFW prompts (Danbooru tags and captions), 4 epochs, batch 32, lr 1e-3, about 14 minutes on my RTX 2060 12GB.</p>

<h3>License</h3>
<p>The head is a Derivative of Anima and is released for <strong>non-commercial use</strong> under the <a href="https://huggingface.co/circlestone-labs/Anima/blob/main/LICENSE.md" rel="ugc">CircleStone Labs Non-Commercial License v1.2</a>. The Qwen3.5-0.8B-Base layers are <a href="https://huggingface.co/Qwen/Qwen3.5-0.8B-Base/blob/main/LICENSE" rel="ugc">Apache-2.0</a> (Copyright Alibaba Cloud), truncated to 20 layers with the final norm replaced by the head. The same notices are in the file's metadata.</p>
<p>The CircleStone Model is licensed by CircleStone Labs LLC under the CircleStone Non-Commercial License. Copyright CircleStone Labs LLC. IN NO EVENT SHALL CIRCLESTONE LABS LLC BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH USE OF THIS MODEL.</p>
<p>Modified by kayfahaarukku: this file is a Derivative of the CircleStone Model, not an official CircleStone release.</p>

<h3>Credits</h3>
<ul>
<li>Qwen team for Qwen3.5-0.8B and Qwen3-0.6B</li>
<li>CircleStone Labs for Anima</li>
<li>Training prompts from <a href="https://huggingface.co/datasets/isek-ai/danbooru-tags-2024" rel="ugc">isek-ai/danbooru-tags-2024</a> (CC0) and <a href="https://huggingface.co/datasets/CaptionEmporium/anime-caption-danbooru-2021-sfw-5m-hq" rel="ugc">CaptionEmporium/anime-caption-danbooru-2021-sfw-5m-hq</a> (CC-BY-SA-4.0)</li>
<li>The head layout (Linear, ExpRMSNorm, SiLU, Linear) follows lylogummy's Qwen3.5-4B Anima encoder</li>
</ul>

<p>Let me know what you think! If you want to try it the easy way, it works out of the box in <a href="https://github.com/nawka12/diffucore-ui" rel="ugc">Diffucore UI</a>.</p>


Qwen3.5-0.8B Text Encoder for Anima

<p>Qwen3.5-0.8B-Base, 20 layers, with a text-distilled projection head. Needs Diffucore v0.2.1+ or the comfyui-qwen35-anima-head node.</p>

Files with SHA256: 555a46ec06b6ec0c4b1f92a663a17c10bce2e0535376ffa69bd7e833366e0474