Qwen4 Training Announced: 10T Params, 33-Round Self-Evolution, and the 4-Tier Strategy

Qwen4 Status: In Training, Not Released

As of September 22, 2026, Qwen4's four tiers — Max, Flash, Plus, and 27B — remain announced but unavailable: no model cards, open weights, API endpoints, pricing, or third-party benchmarks. Alibaba's only timeline signal is "coming soon." The article's value lies in pre-judging Alibaba's 12–24 month competitive moves from the disclosed product matrix and roadmap.

Four-Tier Product Line: Continuity with Qwen3.8

Qwen4-Max : Flagship tier, inheriting Qwen3.8-Max's 2.4 trillion parameters, targeting long-horizon autonomous tasks, research reproduction, and flagship vision/video understanding.

Qwen4-Flash (闪电) : High-speed, low-cost tier, matching Qwen3.8's ¥0.8/¥2.7 extreme cost-performance, intended as the default channel for most API requests.

Qwen4-Plus : Balanced tier between Max and Flash, handling general production workloads that need both capability and cost control.

Qwen4-27B : Open-source tier, explicitly planned for release to developers. Qwen3.8-27B runs on a single 24 GB GPU; Qwen4-27B likely retains this local-inference positioning.

One-line summary: Max is the ceiling, Flash is the cash flow, Plus is the workhorse, 27B is the community engine.

New Architecture: Inherited from Flash-Next, Built for Massive Scale

Qwen4 adopts a next-generation architecture previewed in the open-sourced Qwen3.8-Flash-Next (August 2026). Three core innovations:

Ultra-sparse MoE : 125B total parameters, only 6B activated per token, decoupling model scale from per-step compute cost.

New attention and information-passing design : Block-wise rather than token-wise selection, yielding clear long-context latency advantages.

n-gram Embedding : 51B embedding table offloadable to host memory or NVMe, keeping weights resident without burning GPU memory.

These designs share a single goal: let total parameters grow while keeping single-step inference cost controllable . This architectural foundation makes the 5–10 trillion parameter roadmap viable rather than a pure money-burning exercise.

RSI (Recursive Self-Improvement): 33 Unattended Evolution Rounds

The most striking disclosure is RSI progress:

Qwen3.8-Max completed 33 autonomous iterations in just over a month with zero human intervention , raising its Artificial Analysis intelligence index from 40 to 45 (+12.5%) .

In a chip-design experiment, the model self-improved for 60+ hours, invoked EDA tools 10,000+ times , and produced a manufacturable chip bus module that reduced die area by 42% with zero performance loss .

This demonstrates "model trains model" — autonomous experiment design, data generation, defect discovery, and iteration. Alibaba confirmed RSI will be integrated into Qwen4's training pipeline, drastically cutting manual hyperparameter tuning costs and marking the closest public step toward ASI (Artificial Superintelligence) narratives.

5–10 Trillion Parameter Roadmap: Doubling Down on Scaling

Qwen lead Liu Daiheng stated: "Scaling is a critical path to ASI." The roadmap shows Qwen4.5 and Qwen5 targeting 5–10 trillion parameters , up from Qwen3.8-Max's 2.4 trillion. Alibaba bets that continued model scaling remains the effective frontier lever, not just cost-optimizing mid-size models. Skeptics question inference cost, deployment barriers, and ROI; Alibaba's answer is a "trinity" of chip, cloud, and model co-optimization: proprietary chips lower training/inference cost, Alibaba Cloud provides the compute foundation, and the model layer delivers capabilities.

Full-Stack Foundation: Xuanwu V900 Chip + 20 GW Data Centers

Xuanwu V900 (Zhenwu V900) : T-Head's custom AI training/inference processor — 216 GB VRAM, 1,200 GB/s bandwidth, 3× the performance of the previous M890 , slated for commercial use in early 2027, scalable to 500,000-card super-node clusters .

20 GW capacity target : CEO Wu Yongming set a goal for Alibaba Cloud global data centers to exceed 20 GW by 2032.

Agentic Cloud architecture : AI-native compute layer + AgentCore tooling + Context Engine data layer, cutting token consumption for knowledge-intensive workloads by 67% .

Qwen Intelligence : Mobile cross-app agent framework for handset vendors, embedding agents into hardware.

The subtext: Alibaba avoids a pure model-layer duel, instead competing on total cost of ownership via a vertically integrated chip–cloud–model stack.

Multimodal Suite + Open-Source Ecosystem Expansion

Qwen3.8-LiveTranslate : Simultaneous interpretation model, per-character latency reduced from 2.8 s to 2.3 s .

Qwen-Audio-3.1 series : Full ASR/TTS/Realtime coverage; TTS-Next generates cinematic soundscapes from a text script in one pass.

Qwen-Image 3.1 : Creative design and e-commerce marketing, native transparent backgrounds, image editing, launching within 2026.

Next-gen video generation model : Official preview for November 2026 release.

Open-source metrics: 460+ Qwen models released, 3 billion+ total downloads , 300,000+ derivative models . Qwen3.8 series alone saw 56 million downloads in its first month and spawned 1,900+ derivatives . Qwen4-27B will continue the open-source commitment.

Implications for Developers

Three actionable takeaways:

The downloadable tier remains 27B. Flagship models stay cloud-only; local developers should track 27B, cloud teams should route ~80% of requests to Flash by default, not blindly to Max.

RSI will reshape model iteration cadence. When models run 33 self-improvement rounds and design their own experiments, the marginal value of manual tuning drops rapidly. Future competition shifts to whose data flywheel and validation loop are smoother, not who hires more algorithm engineers.

Scaling returns to the mainstream narrative. If 5–10 trillion parameters materialize, the closed-vs-open gap may widen then compress rapidly. The open-source community must catch up on ultra-large-model runtime, quantization, and inference-engine adaptation — precisely why Alibaba released Flash-Next as an "architecture preview" to let the ecosystem prepare early.

Practical advice: Implement request tiering now — simple QA → Flash, cross-document complex reasoning → Max, sensitive data → local 27B. When Qwen4 ships, you only swap model IDs, not rewrite architecture.

Data sourced from Alibaba 2026 Yunqi Conference official releases, Alibaba Cloud blog, Qwen team on-site materials, and coverage by AIBase and GenAI Daily as of September 24, 2026. Qwen4 tier specifications, pricing, and release dates subject to future official announcements.