<p>This is the GPT-OSS text encoder used by Microsoft's Lens text-to-image model. Lens conditions its 48-block MMDiT denoiser on concatenated multi-layer features from GPT-OSS rather than a CLIP-style text tower, which is a big part of where its long-caption following and multilingual generalization come from. This page mirrors the encoder weights so you can run Lens and Lens-Turbo on-site without pulling them separately.</p>

<p><strong>The underlying model is <a href="https://openai.com/index/introducing-gpt-oss/">GPT-OSS</a>, released by <a href="https://huggingface.co/openai">OpenAI</a> in August 2025 under Apache 2.0.</strong> All credit for the language model goes to OpenAI. The specific bundled weights here come from the <a href="https://huggingface.co/microsoft/Lens">microsoft/Lens</a> repository, where they live under the <code>text_encoder/</code> folder alongside the Lens denoiser and FLUX.2 VAE. Civitai is hosting a mirror so creators can run Lens on-site - head to the upstream repos for the canonical weights and updates.</p>

<h3>Built by</h3>
<p><strong>OpenAI</strong> released the GPT-OSS family (gpt-oss-20b and gpt-oss-120b) in August 2025 as open-weight reasoning models with full chain-of-thought access, MoE architecture, and MXFP4 quantization. <strong>Microsoft Research</strong> selected GPT-OSS as the text encoder for Lens and bundled the weights in their distribution - project leads Dong Chen, Fangyun Wei, and Ziyu Wan, with core contributors Jiawei Zhang, Jinjing Zhao, Sirui Zhang, Yang Yue, and Zhiyang Liang.</p>

<h3><strong>How Lens uses it</strong></h3>
<p>The encoder is frozen - Lens does not finetune it. Features from multiple GPT-OSS layers are concatenated and fed into the MMDiT denoiser as conditioning. This gives Lens stronger long-caption comprehension than a single-final-layer extraction, and it inherits GPT-OSS's multilingual coverage essentially for free.</p>

<h3><strong>Variant</strong></h3>
<p>Microsoft's README does not name the specific GPT-OSS checkpoint, but parameter parity with the 3.8B Lens denoiser strongly points to <strong>gpt-oss-20b</strong> (21B total, 3.6B active) rather than gpt-oss-120b. The bundled weights are stored in MXFP4 by default - inference scripts expose a <code>--disable_mxfp4</code> flag to dequantize when the host GPU does not support it.</p>

<h3><strong>What this model page is for</strong></h3>
<p>You do not generate images with this encoder alone. It is a dependency of Lens and Lens-Turbo, mirrored here so on-site generation can resolve the full pipeline without external fetches. If you are looking for the image generator itself, see the Lens model page.</p>

<h3>Links</h3>
<ul>
  <li>Lens (parent model): <a href="https://huggingface.co/microsoft/Lens">huggingface.co/microsoft/Lens</a></li>
  <li>Lens-Turbo: <a href="https://huggingface.co/microsoft/Lens-Turbo">huggingface.co/microsoft/Lens-Turbo</a></li>
  <li>GPT-OSS (upstream language model): <a href="https://huggingface.co/openai/gpt-oss-20b">huggingface.co/openai/gpt-oss-20b</a></li>
  <li>Lens code: <a href="https://github.com/microsoft/Lens">github.com/microsoft/Lens</a></li>
  <li>Lens license: MIT - GPT-OSS license: Apache 2.0</li>
</ul>

Files with SHA256: 103d7759c720627e5ffdcb0d885595695085dad4201fa6a522a84d4b86335ca0