Sources

Primary first. Recap blogs are not listed unless they were the only place a date showed up.

Qwen3.8

• Qwen3.8-Max launch post (canonical URL; JS-rendered)
• Alibaba Cloud reprint of the same post
• Alibaba launch writeup
• Hugging Face: Qwen3.8-2.4T-A95B
• config.json
• LICENSE
• FP8 sibling
• Alibaba Model Studio: qwen3.8-max

Architecture numbers in this book (layers, heads, experts, hidden size, vocab, native context) come from the model card and config.json. Product claims about 16-day coding runs, chip design, and shop simulations come from the launch blog and are labeled as such.

Lineage

• Qwen3 technical report — training recipe, thinking mode, distillation
• Qwen3-Next: hybrid attention + sparse MoE
• Qwen3.5-397B-A17B model card — same spine, smaller
• Hugging Face: Qwen3.5
• Hugging Face: Qwen3.5 MoE

Systems (part II)

• DeepSeek-V3 technical report — DualPipe, EP, FP8, 2.788M H800-hours, no rollback
• DeepSeek hardware reflections — H800 constraints, disaggregated serving
• ZeRO — 16 bytes/parameter model-state identity
• Megatron-Core parallelism guide
• MoE Parallel Folding
• MegaScale — 12k-GPU MFU, 100+ recoveries
• ByteRobust — failure taxonomy, 97% ETTR
• Kimi K2 — why not DualPipe, QK-Clip, activation offload
• Llama 3 training blog and Meta infra
• Attention sinks · Massive activations
• xAI Colossus — cluster size only; not a training paper
• NVIDIA on Colossus / Spectrum-X

Background

• Attention Is All You Need — the transformer
• GQA
• RoFormer / RoPE
• RMSNorm
• SwiGLU
• BPE
• Gated Delta Networks (the cheap attention cousin)
• Raschka on Gated DeltaNet

What we did not treat as fact

Community Qwen3.8-27B repos that are not under Qwen/. Blog posts that inferred 256 experts (the official number is 512). NVIDIA's writeup where it disagrees with the card on reasoning_effort labels. Any leaderboard number that only exists in Alibaba's table.