Open Source AI Academy
Master the art of Local & Open Source AI. From Transformers and quantization to vLLM production serving, fine-tuning, and training your own models.
The Open Source AI Academy is a free 18-module, 31-lesson curriculum with 31 CLI labs and 73 quizzes into running and fine-tuning open-weights models locally. It covers Meta Llama 4 & 3.3, OpenAI gpt-oss (120B/20B), DeepSeek-V4 & R1, Qwen3/3.8, Mistral Large 3/4, Gemma 4, Kimi K3, vLLM V1 & SGLang serving, Ollama, ggml-org/llama.cpp GGUF quantization, and Unsloth/TRL QLoRA & GRPO alignment.
What is Open Source AI Academy?
Open Source AI Academy is a 100% free, interactive engineering course by Infinity Tech Stack covering self-hosted and open-weight large language models. Across 18 modules, 31 lessons, 31 CLI terminal labs, and 73 quizzes, developers master Llama 4, Mistral, OpenAI gpt-oss, Gemma 4, DeepSeek, Qwen3/3.8, Kimi K3, GGUF/AWQ/FP8 quantization, Ollama, ggml-org/llama.cpp, vLLM V1 & SGLang serving, and Unsloth/TRL LoRA, QLoRA & GRPO post-training with zero sign-up required.
Open Source AI Technical Training Modules
The Open Source AI Space
Transformer Architecture
The Meta Llama Family
The Mistral Ecosystem
DeepSeek & Reasoning
Qwen, Gemma & Others
China Frontier Models
Hugging Face Ecosystem
Quantization Mastery
Ollama: Local AI
llama.cpp Engine
vLLM: Production Serving
SGLang & Alternative Engines
Training & Alignment
Evaluation & Benchmarks
Production MLOps
Applied Local AI & RAG
Frequently Asked Questions
Is the Open Source AI Academy free?
Yes, 100% free. No sign-up, no credit card. All modules, lessons, and quizzes are immediately accessible. Progress saves locally in your browser.
Does this academy cover China's latest frontier open models?
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.
Did Kimi K3 really score higher than Claude?
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.
What is open-source AI?
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 closed-API frontier models (proprietary GPT-5 or Claude tiers), open-weight models like Meta Llama 4, Mistral Large 3/4, DeepSeek-V4/R1, Qwen3, Google Gemma 4, and OpenAI's Apache 2.0 gpt-oss series can be downloaded, run locally, fine-tuned for specific tasks, and deployed without per-token API costs. The ecosystem includes inference engines (ggml-org/llama.cpp, vLLM V1, SGLang, Ollama), post-training stacks (Hugging Face Transformers v5, TRL, Unsloth, Axolotl, PEFT), and quantization tools that make frontier models accessible on local hardware.
What topics does the Open Source AI Academy cover?
The course covers Transformer architecture, Meta Llama 4 & 3.3, Mistral, OpenAI gpt-oss (120B/20B), DeepSeek-V4/R1, Qwen3/3.8, Gemma 4, quantization (GGUF/AWQ/GPTQ/FP8/MXFP4), Ollama, ggml-org/llama.cpp, vLLM V1, SGLang RadixAttention, Unsloth & PEFT LoRA/QLoRA fine-tuning, RLHF/DPO/GRPO alignment, and production MLOps with Docker.
What is quantization?
Quantization reduces a model's numerical precision — typically from 16-bit floating point (FP16/BF16) to 4-bit or 8-bit formats (INT4, NF4, MXFP4, FP8) — to dramatically decrease memory usage and increase inference speed with minimal quality loss. Common formats include GGUF (used by llama.cpp and Ollama, e.g., Q4_K_M, IQ4_XS, Q8_0), AWQ and FP8 (vLLM/SGLang GPU serving), GPTQ, and EXL2. A 70B parameter model that normally requires 140GB of VRAM in FP16 can run in ~35–40GB with 4-bit quantization.
Do I need programming experience?
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.
What is vLLM?
vLLM is the industry-standard open-source engine for high-throughput production LLM serving. Its V1 engine architecture pairs PagedAttention — managing GPU KV cache like OS virtual memory pages — with zero-overhead Automatic Prefix Caching, continuous batching, tensor/expert parallelism, FP8/AWQ quantization, and OpenAI-compatible API endpoints with native tool calling (--enable-auto-tool-choice).