Keyvan.ai 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.
Current research: weight-space programming
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 Sparks of Weight-Space Programming (2 November 2026), we replicated and extended the approach with Claude Mythos 5.1.
- Mythos-WSP-Qwen3.5-0.8B – 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.
- Mythos-WSP-Loopbreaker-Skill – the implanted circuit as a separate, versioned tensor set with an application script.
- Research report – Weight-Space Programming with Claude Mythos (bilingual EN/DE), DOI 10.5281/zenodo.24036187, including a transfer of the method to a thermal surrogate model from power electronics.
Models
German language (Mankei family)
- Mankei-1B-Chat · Mankei-1B-Chat-GGUF – 1B German chat model, safetensors and GGUF
- Mankei-Boarisch – 1B Bavarian-dialect text generation
- Mankei-326M-Embedder · Mankei-326M-Reranker – German retrieval: sentence embeddings and reranking
- Mankei-Reflex – text classification
- german-text-3.1 – 2B German text generation
German document understanding (German-OCR family)
- german-ocr-3.1 (2B) · german-ocr-3 (0.8B) – image-text-to-text models for German documents
- german-ocr · german-ocr-2b-gguf · german-ocr-turbo · g-ocr – earlier generations and GGUF builds
Vision and industrial inspection
- belegant-4b – 4B image-text-to-text model for receipts and business documents
- AnomalyDetection-MVTech-Metal – image segmentation for surface anomalies on metal parts
Audio
- wakecore – wake-word and audio classification
Security: intrusion detection for vehicle and IT networks
- CanDefender · CanDefender-Fuzzy · CANDefender-DoS – CAN-bus intrusion detection (fuzzing and DoS attacks)
- SecIDS-v2 · SecIDS-CNN – network intrusion detection
Principles
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.