Is the Open Source AI Academy free?

Yes, 100% free. No sign-up, no credit card. All 18 modules, 31 lessons, and 73 quiz questions are accessible immediately. Your progress saves locally in your browser.

How is Infinity Open Source AI Academy different from vendor-specific open-weight tutorials?

Unlike generic or single-model tutorials, Infinity Tech Stack's Open Source AI Academy bridges global open weights (Meta Llama 4 & 3.3, Mistral Large 3/4, Google Gemma 4, OpenAI gpt-oss-120b/20b) and China Frontier Models (DeepSeek-V4 & R1, Qwen3/3.8, Kimi K3, GLM-5) with 18 comprehensive modules and simulated in-browser terminal labs for Ollama, ggml-org/llama.cpp, vLLM V1, SGLang, Unsloth, and GGUF quantization. It requires zero GPU hardware, no cloud subscriptions, and no sign-up.

Does the Open Source AI Academy cover China's latest frontier open models?

Yes. The academy includes a dedicated China Frontier Models module covering Qwen, DeepSeek, Z.ai GLM, Moonshot Kimi, InternLM, Baichuan, and Yi, plus a practical model-radar workflow for weekly tracking and gated production adoption.

Did Kimi K3 score higher than Claude?

On several benchmarks, yes. Kimi K3 (July 14, 2026: 2.8T open-weight MoE, 1M context) beat Claude Fable 5 on 6 of 14 shared benchmarks — including BrowseComp (91.2 vs 88.0), SWE Marathon (42.0 vs 35.0), and Terminal-Bench 2.1 (88.3 vs 84.6) — topped the Arena.ai Frontend Code Arena, and edged Claude Opus 4.8 on the Artificial Analysis Intelligence Index (57 vs 56) at about 3.3x lower pricing. Claude Fable 5 won the other 8 benchmarks and leads that index at 60. The academy's China Frontier Models module has a dedicated lesson on the full scoreboard.

What topics does the Open Source AI Academy cover?

The course covers Transformer architecture, Meta Llama 4, Mistral Large 3/4, OpenAI gpt-oss (120B/20B), DeepSeek-V4/R1, Qwen3/3.8, Google Gemma 4, Hugging Face Transformers v5, quantization (GGUF, AWQ, GPTQ, FP8, MXFP4, EXL2), Ollama, ggml-org/llama.cpp, vLLM V1, SGLang RadixAttention, Unsloth & PEFT LoRA/QLoRA fine-tuning, pre-training from scratch, and TRL RLHF/DPO/GRPO alignment, plus production MLOps with Docker and Kubernetes.

vLLM is the industry-standard open-source engine for high-throughput production LLM serving. Its V1 engine architecture pairs PagedAttention virtual memory paging with zero-overhead Automatic Prefix Caching, continuous batching, FP8/AWQ quantization, and native tool calling (--enable-auto-tool-choice) for >90% GPU utilization.

Do I need expensive hardware to run open-source models?

No. Many models can run on consumer GPUs (e.g., RTX 4090) or even CPUs via llama.cpp. The course covers deployment strategies for every budget, from Raspberry Pi edge devices to multi-GPU production clusters.

Does the Open Source AI Academy include hands-on labs?

Yes. Hands-on terminal labs are built into the lessons - ollama pull, run, and list models, quantize a model to Q4_K_M, and serve it with vLLM using real, vendor-verified commands, with objectives, hints, solutions, and XP rewards. No GPU or sign-up needed.

Which Google Gemma 3n model variant (~2B effective parameters) is designed for smartphones and IoT?

Gemma 3n E2B — Gemma 3n E2B (~2B effective parameters via MatFormer Per-Layer Embeddings) is optimized for resource-constrained edge devices like smartphones and IoT hardware.

The Global Model Families — Qwen, Gemma & Others

[ ABORT TO HUD ]

SEQ. 1

The Global Model Families

🌏 Qwen, Gemma & Others⏱ 10 min⭐ 100 BASE XP⌨ HANDS-ON LAB

Beyond Llama & Mistral

Qwen (Alibaba)

Alibaba's Qwen3, Qwen3-Coder, Qwen 3.5 (hybrid Gated DeltaNet + sparse MoE), and Qwen 3.8 (August 2026) anchor the Apache 2.0 open-weight ecosystem alongside specialized omnimodal and robotics releases.

Model	Type	Active Params	Best For
Qwen3.8-Max	API-Only	2.4T	Proprietary flagship (Aug 3, 2026)
Qwen3.8-2.4T-A95B	MoE	95B of 2.4T	Open-weights frontier MoE (Aug 12, 2026)
Qwen3.8-27B	Dense	27B	Dense multimodal, Apache 2.0 (Aug 14, 2026)
Qwen3.6-27B	Dense	27B	Consistent high performance (legacy)
Qwen3.6-35B-A3B	MoE	3B of 35B	Ultra-efficient inference (legacy)
Qwen3-Coder-480B-A35B	MoE	35B of 480B	Apache 2.0 repository-scale coding (256K native context)
Qwen3-Coder-Next	Hybrid MoE	3B of 80B	Apache 2.0 ultra-fast local coding agent
Qwen3-235B-A22B & 30B-A3B	MoE	22B / 3B active	Apache 2.0 general reasoning & single-GPU workhorses

Key feature: Thinking/Non-Thinking modes (introduced in Qwen 3 and expanded in Qwen 3.6/3.8) - switch between deep reasoning and fast responses in a single unified model.

Google Gemma 4 (Apache 2.0) & Gemma 3n

In April 2026, Google released Gemma 4 under the fully permissive Apache 2.0 license (upgrading from Gemma 3's custom Gemma Terms of Use), pairing edge MatFormer models with server MoE and dense checkpoints:

Gemma 4 E2B & Gemma 3n E2B: ~2B effective params (MatFormer Per-Layer Embeddings) with native audio/image/text for smartphones & IoT
Gemma 4 E4B & Gemma 3n E4B: ~4B effective params for flagship mobile, laptops & edge devices (128K context)
Gemma 4 26B-A4B (MoE) & 31B (Dense): Apache 2.0 workstation/server models activating ~3.8B–4B (MoE) or 31B (Dense) with 256K context, native vision/video, thinking mode, and built-in function calling (succeeding Gemma 3 1B/4B/12B/27B)
Runs on Android, iOS, Raspberry Pi, Ollama, llama.cpp, vLLM, and WebGPU / MediaPipe

OpenAI `gpt-oss` (Apache 2.0 Open Weights)

OpenAI entered the permissive open-weight ecosystem with two native Mixture-of-Experts reasoning models under Apache 2.0, post-trained with native MXFP4 quantization, 131K context, full Chain-of-Thought inspection, native tool calling, and configurable reasoning effort (low, medium, high):

Model	Total / Active Params	Hardware Footprint	Highlights
`gpt-oss-120b`	~117B total / 5.1B active (128 experts, top-4)	Single 80GB GPU (H100 / MI300X in MXFP4)	Frontier open reasoning, coding & agentic tool use
`gpt-oss-20b`	~21B total / 3.6B active	16GB RAM/VRAM (consumer GPUs & laptops)	Low-latency local reasoning & structured outputs

Other Notable Families

Family	Creator	Standout Feature
Phi-4	Microsoft	Small but mighty (14B rivals 70B models)
Command-R+	Cohere	Optimized for RAG & enterprise search
Yi-Lightning	01.AI	Chinese-English bilingual excellence

🐳 Edge Container: Run Gemma 3n in Docker with Ollama:
docker run -d --gpus all -p 11434:11434 ollama/ollama && docker exec -it $(docker ps -q) ollama run gemma3n:e4b

⌨ HANDS-ON LABBenchmark Qwen 2.5 Coder Output

⭐ +150 XP

Run Qwen 2.5 Coder in the terminal to generate an optimized binary search function.

1Invoke qwen2.5-coder via ollama with a code generation prompt.

lab-sandbox — simulated environment

INFINITY LAB SANDBOX v2.6 — simulated shell
Type the command for the current objective. Helpers: "hint", "solution", "clear".
$

OBJECTIVE 1 / 1 — type "hint" if stuck

KNOWLEDGE CHECK

QUERY 1 // 2

What is unique about Qwen 3 / 3.6's Thinking/Non-Thinking modes?

Different model files

Toggleable deep reasoning vs fast responses in one model

Only works in Chinese

Requires special hardware