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 (Llama 3.3, Mistral) and China Frontier Models (DeepSeek R1, Qwen 2.5, Kimi K3, GLM) with 18 comprehensive modules and simulated in-browser terminal labs for Ollama, llama.cpp, vLLM, 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, DeepSeek, Qwen, Gemma, Hugging Face, quantization (GGUF, AWQ, GPTQ, EXL2), Ollama, llama.cpp, vLLM, SGLang, LoRA/QLoRA fine-tuning, pre-training from scratch, RLHF/DPO/GRPO alignment, and production MLOps with Docker and Kubernetes.

vLLM is the industry-standard engine for high-throughput production LLM serving. It uses PagedAttention for efficient GPU memory management and continuous batching 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 two kernel backends does Axolotl v0.18 add for NVFP4 MoE-LoRA?

ScatterMoE and SonicMoE — ScatterMoE (fused NVFP4/bnb expert paths for Gemma 4 and DeepSeek-V3) and SonicMoE (native FP4-activation W4A4 MoE-LoRA via CUTLASS, for Qwen3-30B-A3B and Qwen3-Next-80B-A3B) are the two backends.

What does the new hidden_states activation offloading mode enable on Qwen3-8B full-parameter training?

128K context where plain gradient checkpointing OOMs — By offloading only the per-layer checkpoint input to CPU and recomputing the rest, Qwen3-8B full-parameter training reaches 128K context with bit-exact gradients, where plain gradient checkpointing would run out of memory.

Axolotl v0.18: Production-Scale MoE Fine-Tuning Tutorial | Fine-Tuning & LoRA

[ ABORT TO HUD ]

SEQ. 1

SEQ. 2

Axolotl v0.18: Production-Scale MoE Fine-Tuning

🎯 Fine-Tuning & LoRA⏱ 13 min⭐ 150 BASE XP

Fine-Tuning Frontier MoE Models Cheaply

Axolotl v0.18 (released July 2026) is built around one theme: making 4-bit expert LoRA on very large sparse Mixture-of-Experts models fast and memory-flat at long context, on both Blackwell and Hopper GPUs.

NVFP4 MoE-LoRA: Two Kernel Backends

Backend	Approach	Verified On
ScatterMoE	Fused NVFP4 (Marlin/DeepGEMM) and bnb-4bit expert paths	Gemma 4 (128 experts), DeepSeek-V3
SonicMoE	Native FP4-activation MoE-LoRA (W4A4) via quack/CUTLASS kernels - beats Marlin W4A16 at every sequence length on B200	Qwen3-30B-A3B, Qwen3-Next-80B-A3B

Active VRAM stays nearly flat from 4K to 32K context with both backends, and SonicMoE now also runs on consumer Blackwell (RTX 50-series) via a quack 0.6 migration.

New Model Support

GLM-5.2 (glm_moe_dsa): fine-tune from its NVFP4 checkpoint on multi-GPU FSDP2 - 256 routed experts with Lightning-Indexer token selection, following the DeepSeek-V3.2 sparse-MLA lineage, with DeepEP expert parallelism composed in 2D (EP x cp)
Gemma 4 Unified: encoder-free multimodal text + vision LoRA configs with a dedicated chat template
DeepSeek-V3: MoE-LoRA kernels alongside Gemma 4
MiniMax M2 (MoE): QLoRA fine-tuning with kernel-optimization patches
PaddleOCR-VL: multimodal image-text support

Long-Context Training Without OOMs

A new activation_offloading: hidden_states mode offloads only the per-layer checkpoint input to CPU and recomputes the rest - on Qwen3-8B full-parameter training it reaches 128K context where plain gradient checkpointing runs out of memory, with bit-exact gradients.

Multi-Turn Inference Chat Interface

axolotl inference config.yaml --chat

This starts a streaming, multi-turn chat session with runtime-adjustable generation parameters and streaming reasoning blocks for thinking models - useful for sanity-checking a fine-tune without writing a separate eval script.

Dependency Bumps

Package	Change
transformers	5.9.0 → 5.14.1
trl	1.5.1 → 1.8.0
liger-kernel	0.7.0 → 0.8.0
huggingface_hub	1.1.7 → 1.17.0

💡 Why This Matters: Fine-tuning frontier-scale sparse MoE models (GLM-5.2, DeepSeek-V3, Qwen3-Next-80B-A3B) used to require full-precision experts and multi-node clusters. NVFP4 MoE-LoRA collapses that to single-node, memory-flat training - the same democratization LoRA brought to dense models now extends to the largest open MoE architectures.

KNOWLEDGE CHECK

QUERY 1 // 3

What is the primary theme of Axolotl v0.18?

Adding image-generation support

Making 4-bit expert LoRA fast and memory-flat on large sparse MoE models

Removing GPU support entirely

Replacing PEFT with a new library