Weekend hack: Project Titania reimplements Qwen3-0.6B from transformer to GPU ISA simulator
generativist · x · 2026-09-14
Developer penberg open-sourced Titania, a from-scratch LLM stack where every layer — model kernels, instruction set, compiler, and GPU simulator — is small enough for one person to read and implement. It runs a real Qwen3-0.6B decoder-only transformer you can chat with; kernels compile to the Titania ISA and execute on an ISA simulator, with RTL GPU, FPGA, and silicon on the roadmap. Built as a weekend hack with help from Claude Code, it demystifies how LLMs actually run on massively parallel GPU hardware.
More from coding & agent
- Agent observability is a distraction: tool-call success isn't outcome correctness — Gallegos_Daniel · 2026-09-14
- Agent engineering lessons: the model is the smallest, most swappable part — alex_verem · 2026-09-14
- Running Qwen 27B on 2x P40s: parallel agents wreck KV cache, seeking a serial agent harness — Jumpy-Operation-4615 · 2026-09-14
- Runway launches MCP: generate video directly in ChatGPT, Claude and Cursor — tlakomy · 2026-09-14
- Google open-sources ARTEMIS, an Android automation agent — kunalpuri · 2026-09-14
- Testing Supermemory, Mem0, Vilix: cloud memory tools share one structural flaw — Asly97 · 2026-09-14