| <!doctype html><html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1"><title>Keyvan.ai</title> |
| <style>html,body{margin:0;background:#ffffff;color:#1a1a1a;font:16px/1.6 system-ui,-apple-system,"Segoe UI",sans-serif}main{max-width:960px;margin:0 auto;padding:32px 24px 64px}img{max-width:100%;border-radius:10px;display:block;margin:8px 0 16px} |
| h2{font-size:20px;margin:30px 0 8px;color:#111111}strong{color:#111111;font-weight:700}a{color:#0b5fa5;text-decoration:none}a:hover{text-decoration:underline}ul{padding-left:20px;margin:4px 0 10px}li{margin:3px 0}p{margin:8px 0}</style></head><body><main> |
| <img src="https://huggingface.co/spaces/keyvan-ai/README/resolve/main/banner.png" alt="Keyvan.ai"> |
| <p><strong>Keyvan.ai</strong> is an independent AI research and engineering practice based in Germany (Keyvan Hardani). We build and evaluate language, vision, audio and security models for production use, with a focus on small models that run on-premises, and we publish our experiments with full protocols.</p> |
| <h2>Current research: weight-space programming</h2> |
| <p>Editing the weights of neural networks directly with a frontier model as the agent, without gradients or training, verified by a fixed evaluation harness. Following the MIT CSAIL study <em>Sparks of Weight-Space Programming</em> (2 November 2026), we replicated and extended the approach with Claude Mythos 5.1.</p> |
| <ul> |
| <li><a href="https://huggingface.co/keyvan-ai/Mythos-WSP-Qwen3.5-0.8B">Mythos-WSP-Qwen3.5-0.8B</a> – Qwen3.5-0.8B with two weight edits: a rank-3 projection and an implanted repetition-detection circuit in four repurposed MLP units. MATH-500 (thinking mode, 32k tokens): 7.3 % → 58.7 %, truncated answers 93 % → 17 %. No training.</li> |
| <li><a href="https://huggingface.co/keyvan-ai/Mythos-WSP-Loopbreaker-Skill">Mythos-WSP-Loopbreaker-Skill</a> – the implanted circuit as a separate, versioned tensor set with an application script.</li> |
| <li><a href="https://doi.org/10.5281/zenodo.24036187">Research report</a> – <em>Weight-Space Programming with Claude Mythos</em> (bilingual EN/DE), DOI 10.5281/zenodo.24036187, including a transfer of the method to a thermal surrogate model from power electronics.</li> |
| </ul> |
| <img src="https://huggingface.co/spaces/keyvan-ai/README/resolve/main/abb1_math500.png" alt="MATH-500"> |
| <h2>Models</h2> |
| <p><strong>German language (Mankei family)</strong></p> |
| <ul> |
| <li><a href="https://huggingface.co/keyvan-ai/Mankei-1B-Chat">Mankei-1B-Chat</a> · <a href="https://huggingface.co/keyvan-ai/Mankei-1B-Chat-GGUF">Mankei-1B-Chat-GGUF</a> – 1B German chat model, safetensors and GGUF</li> |
| <li><a href="https://huggingface.co/keyvan-ai/Mankei-Boarisch">Mankei-Boarisch</a> – 1B Bavarian-dialect text generation</li> |
| <li><a href="https://huggingface.co/keyvan-ai/Mankei-326M-Embedder">Mankei-326M-Embedder</a> · <a href="https://huggingface.co/keyvan-ai/Mankei-326M-Reranker">Mankei-326M-Reranker</a> – German retrieval: sentence embeddings and reranking</li> |
| <li><a href="https://huggingface.co/keyvan-ai/Mankei-Reflex">Mankei-Reflex</a> – text classification</li> |
| <li><a href="https://huggingface.co/keyvan-ai/german-text-3.1">german-text-3.1</a> – 2B German text generation</li> |
| </ul> |
| <p><strong>German document understanding (German-OCR family)</strong></p> |
| <ul> |
| <li><a href="https://huggingface.co/keyvan-ai/german-ocr-3.1">german-ocr-3.1</a> (2B) · <a href="https://huggingface.co/keyvan-ai/german-ocr-3">german-ocr-3</a> (0.8B) – image-text-to-text models for German documents</li> |
| <li><a href="https://huggingface.co/keyvan-ai/german-ocr">german-ocr</a> · <a href="https://huggingface.co/keyvan-ai/german-ocr-2b-gguf">german-ocr-2b-gguf</a> · <a href="https://huggingface.co/keyvan-ai/german-ocr-turbo">german-ocr-turbo</a> · <a href="https://huggingface.co/keyvan-ai/g-ocr">g-ocr</a> – earlier generations and GGUF builds</li> |
| </ul> |
| <p><strong>Vision and industrial inspection</strong></p> |
| <ul> |
| <li><a href="https://huggingface.co/keyvan-ai/belegant-4b">belegant-4b</a> – 4B image-text-to-text model for receipts and business documents</li> |
| <li><a href="https://huggingface.co/keyvan-ai/AnomalyDetection-MVTech-Metal">AnomalyDetection-MVTech-Metal</a> – image segmentation for surface anomalies on metal parts</li> |
| </ul> |
| <p><strong>Audio</strong></p> |
| <ul> |
| <li><a href="https://huggingface.co/keyvan-ai/wakecore">wakecore</a> – wake-word and audio classification</li> |
| </ul> |
| <p><strong>Security: intrusion detection for vehicle and IT networks</strong></p> |
| <ul> |
| <li><a href="https://huggingface.co/keyvan-ai/CanDefender">CanDefender</a> · <a href="https://huggingface.co/keyvan-ai/CanDefender-Fuzzy">CanDefender-Fuzzy</a> · <a href="https://huggingface.co/keyvan-ai/CANDefender-DoS">CANDefender-DoS</a> – CAN-bus intrusion detection (fuzzing and DoS attacks)</li> |
| <li><a href="https://huggingface.co/keyvan-ai/SecIDS-v2">SecIDS-v2</a> · <a href="https://huggingface.co/keyvan-ai/SecIDS-CNN">SecIDS-CNN</a> – network intrusion detection</li> |
| </ul> |
| <h2>Principles</h2> |
| <p>Every published result comes with the protocol it was measured under, the controls that were run, and the limitations we found. Models are released under their base-model licences; reports under CC BY 4.0.</p> |
| <p><a href="https://keyvan.ai">keyvan.ai</a> · <a href="https://github.com/keyvanhardani">GitHub</a> · <a href="https://www.linkedin.com/in/keyvanhardani">LinkedIn</a></p> |
| </main></body></html> |
