Implement Qwen3.8‑27B from scratch in PyTorch.
Part 1 is free: how a frontier LLM turns text into tokens and predicts the next one.
Free · 19 min · plays right here once you enter your email.
Get Part 1 free →
You'll get Part 1 now, plus a few emails about the full course. Unsubscribe in one click.
You're in. Part 1 is ready — press play.We've also emailed you the link.
After Part 1, the full course:
See the full course →
- The building blocks of Qwen and today's open LLMs
- Gated DeltaNet: a fixed-size memory instead of a growing cache
- Assemble Qwen3.8‑27B and load the official weights
- Inference, then DeepSeek 7B & 67B from the same building blocks
In Part 1
- What a language model actually does — one job, done over and over
- What a token is, and why models don't see words
- How a tokenizer is trained: byte-pair encoding in 4 steps
- Why "Hello" and "hello" are different tokens
Where Part 1 leads
- The building blocks of Qwen and today's open LLMs
- Gated DeltaNet: a fixed-size memory instead of a growing cache
- Assemble Qwen3.8‑27B and load the official weights
- Inference, then DeepSeek 7B & 67B from the same building blocks
4,000+ students · 2,900+ GitHub stars