Qwen (Alibaba)

The leading Chinese open-weight family the synthesis kept naming (cost-per-quality leader on llm-benchmarks) but had no page for. Built by Alibaba Cloud (a.k.a. Tongyi Qianwen); first launched 2023, now at Qwen3 (April 2025) with rapid point releases (Qwen3.5 Feb 2026, Qwen3.6 Apr 2026, and the Qwen3.7/3.8 series through mid-2026). The clearest single rival to deepseek for the open-weight crown, and a heavyweight on the open-weight-models axis.

Why it matters to the market thesis

• Open-weight breadth. Dense variants 0.6B → 32B, plus MoE (e.g. 30B-A3B — 30B total, 3B activated), many under Apache-2.0 — the clean-licensing pole the synthesis contrasts with llama's custom non-OSI license. So Qwen sits with gemma-4/mistral-ai on permissive licensing, but unlike Google it is open-first, not dual-track.
• Full-modality ladder. Reasoning (QwQ-32B), vision-language (Qwen-VL at 3/7/32/72B), audio, Coder specialists, and omni-modal (text+image+video+audio in) — the same "one pitch across the whole stack" pattern the synthesis noted for Google, but shipped as open weights.
• Scale of adoption. 200,000+ Qwen derivatives on Hugging Face (the open-weight ecosystem-seeding effect made concrete) and 234M app users by May 2026. Trained on 36T tokens / 119 languages.

Shipping the next architecture before the next family

Qwen has a habit worth naming as a strategy: it releases the structural changes of the next generation as open weights, ahead of the family built on them. Qwen3-Next did this for Qwen3.5 — its hybrid Gated DeltaNet + gated attention design then carried through Qwen3.5, 3.6, 3.7 and 3.8 — and Qwen3.8-Flash-Next (2026-08-26) does it again, described in its own launch post as "an early preview of the architecture used in Qwen4." No other lab in this wiki lets the community examine its next flagship's geometry a generation early, and given the derivative ecosystem below, that is not only openness — it seeds the shape.

The architecture is becoming infrastructure

The 200,000+ derivatives number understates the reach, because the architecture travels separately from the weights. Two entries in this wiki are built on Qwen3's geometry and disclose it very differently: agents-a1 names its "Qwen3 lineage" in its own materials, while neutrino-1 states a configuration that matches Qwen3-8B/config.json field for field — 36 layers, 4,096 hidden, 32 query over 8 KV heads, 128-dim heads, 12,288 feed-forward, 151,936 vocabulary, 40,960 positions, untied embeddings — and never mentions Qwen.

Nothing improper follows from either. Configurations are not proprietary and Apache-2.0 permits the rest. But it is a real market fact for this spoke: Qwen's contribution to the open ecosystem is increasingly a default shape as much as a set of checkpoints, and a shape leaves no attribution trail. Where a downstream lab is silent about it, this wiki records the match and not a verdict.

qwen38-flash-next extends the same point in a new direction: 125B total, 6B active, plus 51B of N-gram embedding parameters that never enter the per-token matrix-multiply budget, served at $0.15/$0.47 per 1M. Capacity, compute and price are three separate numbers here, and the pricing is budget-tier against a benchmark set Qwen picks to include Claude Opus 4.6.

Where it sits

Qwen sharpens the synthesis's central tension (collapsing floor vs premium frontier): it is the open-weight challenger with the deepest model ladder + cleanest license + biggest derivative ecosystem, the strongest evidence that open weights are closing on the proprietary frontier from below. Its geopolitics/licensing-acceptance question is the same open one flagged for deepseek (Chinese-lab enterprise/regulatory acceptance is unmodeled here).

Related

deepseek · open-weight-models · llm-benchmarks · gemma-4 · mistral-ai · llama · agents-a1 · neutrino-1 · llm-provider · qwen38-flash-next · qwen38-flash-next-willison