Qwen - Miles

Miles ships ready-to-run RL recipes for every generation of the Qwen line: the dense Qwen3 series (0.6 B → 32 B), the Qwen3.5 family with its gated-attention architecture, the Qwen3 and Qwen3.5 MoE variants, and the Gated-Delta-Net Qwen3-Next-80B-A3B.

Variants

Family	Class	Sizes	Recipe
Qwen3	Dense	0.6 B · 1.7 B · 4 B · 8 B · 14 B · 32 B	qwen3
Qwen3	MoE	3 B / 30 B · 22 B / 235 B	qwen3-moe
Qwen3.5	Dense	4 B · 9 B · 27 B	qwen3-5
Qwen3.5	MoE	3 B / 35 B	qwen3-5-moe
Qwen3-Next	MoE (GDN)	3 B / 80 B	qwen3-next

Fastest path to train

Qwen3-4B on a single 8× H100 node — the canonical starter recipe:

cd /root/miles
hf download Qwen/Qwen3-4B --local-dir /root/Qwen3-4B
bash scripts/run-qwen3-4B.sh

Dataset is DAPO-Math-17k at /root/dapo-math-17k/dapo-math-17k.jsonl. See the Qwen3 Dense page for the full walkthrough, weight conversion, and variants.

Which variant do I pick?

Learning Miles for the first time → Qwen3-4B (qwen3). Fits one H100 node, fast loop.
Need MoE on a single node → Qwen3-30B-A3B (qwen3-moe).
Scaling to multi-node → Qwen3-235B-A22B (qwen3-moe).
Latest dense architecture (gated attention, A_log FP32) → Qwen3.5-4B (qwen3-5).
Hybrid MTP / speculative decoding experiments → Qwen3.5-35B-A3B (qwen3-5-moe).
Gated-Delta-Net (fla backend, real-CP) → Qwen3-Next-80B-A3B (qwen3-next).

Inkling-Small Qwen3.6