Open Source AI Academy

Yes, 100% free. No sign-up, no credit card. All modules, lessons, and quizzes are immediately accessible. Progress saves locally in your browser.

Yes. The curriculum now includes a dedicated China Frontier Models track covering Qwen, DeepSeek, Z.ai/GLM, Moonshot/Kimi, InternLM, Baichuan, and Yi — with a practical model-radar workflow for weekly release tracking, evaluation, and production promotion.

On several benchmarks, yes. Kimi K3 (released July 14, 2026: 2.8T-parameter open-weight MoE, 1M-token context, native vision) beat Claude Fable 5 on 6 of 14 shared benchmarks at launch — including BrowseComp web research (91.2 vs 88.0), SWE Marathon long-horizon coding (42.0 vs 35.0), and Terminal-Bench 2.1 (88.3 vs 84.6) — took #1 on the Arena.ai Frontend Code Arena, and edged Claude Opus 4.8 on the Artificial Analysis Intelligence Index (57 vs 56), all at roughly 3.3x lower API pricing. Claude Fable 5 kept the overall lead, winning the other 8 benchmarks and scoring 60 on that index. The China Frontier Models track includes a dedicated lesson with the full scoreboard and the caveats.

Open-source AI refers to artificial intelligence models, frameworks, and tools whose weights, code, or training data are publicly available for inspection, modification, and redistribution. Unlike proprietary models (GPT, Claude), open-weight models like Llama, Mistral, and DeepSeek can be downloaded, run locally, fine-tuned for specific tasks, and deployed without per-token API costs. The open-source AI ecosystem includes inference engines (llama.cpp, vLLM, Ollama), training frameworks (Hugging Face Transformers, Axolotl), and quantization tools that make large models accessible on consumer hardware.

The course covers Transformer architecture, Llama 4, Mistral, DeepSeek, quantization (GGUF/AWQ/GPTQ), Ollama, llama.cpp, vLLM, SGLang, LoRA fine-tuning, RLHF/DPO alignment, and production MLOps with Docker.

Quantization reduces a model's numerical precision — typically from 16-bit floating point (FP16) to 4-bit or 8-bit integers (INT4/INT8) — to dramatically decrease memory usage and increase inference speed with minimal quality loss. Common formats include GGUF (used by llama.cpp and Ollama), GPTQ (GPU-optimized), AWQ (activation-aware), and EXL2 (variable bit-width). A 70B parameter model that normally requires 140GB of VRAM in FP16 can run in ~35GB with 4-bit quantization, making it feasible on consumer GPUs.

Basic familiarity with Python or JavaScript is helpful but not required. The course starts from absolute fundamentals and progressively advances to expert topics like multi-agent orchestration and production deployment.

vLLM is the industry-standard open-source engine for high-throughput production LLM serving. Its core innovation, PagedAttention, manages GPU memory like an operating system's virtual memory — dynamically allocating and deallocating KV cache blocks to serve many concurrent requests efficiently. vLLM supports continuous batching, tensor parallelism for multi-GPU inference, speculative decoding, and OpenAI-compatible API endpoints. It powers production deployments at companies serving millions of requests per day.