Alibaba Unveils 5-10 Trillion Parameter Model Plan and Zhenwu V900 Chip at Apsara Conference

At Apsara Conference 2026 in Hangzhou, Alibaba CEO Eddie Wu unveiled an aggressive full-stack AI strategy: per Reuters, Alibaba plans to train models with 5-10 trillion parameters, while T-Head launched the Zhenwu V900, billed as China's most powerful AI chip. The dual bet on silicon and scale marks a major move in the frontier AI race.

Confirmed

• Wu claimed "machine thinking" today represents less than 3% of human capacity, with room to scale 1000x, as machines become the primary driver of thinking
• Model track: official push on recursive self-improvement (RSI), letting models find their weaknesses from real task feedback, design experiments, and generate training data; per @lxfater, Qwen3.8-Max has run 33 rounds of this loop for over a month; per @teortaxesTex, Wu said the Qwen team is making progress on RSI and expects inference data consumption to rise sharply
• Per Reuters, Alibaba plans a 5-10 trillion parameter model; community leaks (@ChrisGPT, @markk) suggest this scale is slated for Qwen 4.5 and Qwen 5, with Qwen 4 already in training (@ChrisGPT notes Alibaba's Aug 26 architecture preview already contained qwen4exp code); Wu also previewed the next-generation Qwen; @julianharris reports Qwen-Image-3.1 is due by year-end
• Chip track: T-Head's Zhenwu V900 delivers 3x the compute of the M890 released four months prior, with 216GB memory and cluster scaling to 500K cards; positioned against Nvidia as a key step toward a vertically integrated chip-to-datacenter AI stack
• Per @bookwormengr, the 2.4T-parameter Qwen 3.8 Max reportedly runs inference entirely on domestic infrastructure using the Zhenwu M890 (144GB memory, 800GB bandwidth) and Pangu supernodes — a first for a frontier-scale MoE workload on China's homegrown stack
• The conference released a dozen-plus products across three pillars: AI models, AI chips, and AI cloud (@智东西)
• Citing Kyle Chan: Alibaba plans 20GW of global compute by 2032; Zhenwu V900 enters mass production in Q1 next year
• Multiple posts note ASI (artificial superintelligence) as Alibaba's long-term goal

Unconfirmed

• Full Zhenwu V900 specs, interconnect details, and production timeline remain undisclosed; the production date comes only from relayed leaks
• Timing, compute sources, and final form of the 5-10T parameter model remain at planning/leak stage; RSI technical details also come mostly via community relays
• The 20GW datacenter plan, Qwen 4 training progress, and Qwen 4.5/5 parameter allocation all come from unofficial channels pending confirmation

Why it matters

If realized, 5-10 trillion parameters would far exceed current frontier models (Qwen 3.8 is 2.4T, implying a 2-4x jump), showing Alibaba pushing simultaneously on compute (homegrown chips) and models (mega-scale training). Qwen 3.8 Max running entirely on domestic silicon marks the first time China's chip stack serves frontier-scale inference; 500K-card cluster scaling and the 20GW plan reflect a new approach to breaking training compute bottlenecks, warranting close attention to official disclosures and actual training progress.

2026-09-22 ~ 2026-09-23 · 27 related posts

• [source] Alibaba Plans AI Model With 5-10 Trillion Parameters, Unveils New Chip — tengo_harambe · 2026-09-22
• Alibaba roadmap: Qwen explores RSI, new models up to 10T params, 20GW datacenters by 2032 — teortaxesTex · 2026-09-22
• Alibaba unveils new AI chip it calls China's most powerful — AIFlow_ML · 2026-09-22
• Alibaba reported targeting 5-10T-parameter Qwen models, training Qwen 4 with 20 GW datacenter plan — mark_k · 2026-09-22
• Alibaba Unveils Zhenwu V900, China's Most Powerful AI Chip, Scaling to 500K Cards — CurieuxExplorer · 2026-09-22
• Alibaba bets full-stack on AI: 5-10T param Qwen, Zhenwu V900 chip, 20GW cloud by 2032 — Scobleizer · 2026-09-22
• Alibaba CEO Eddie Wu says new AI model to train on 5-10 trillion parameters — JosephJacks_ · 2026-09-22
• Alibaba lays out AI roadmap: 10T-parameter Qwen, Zhenwu V900 chip, 20GW data centers by 2032 — iamfakhrealam · 2026-09-22
• Alibaba CEO Eddie Wu: Machine thinking is under 3% of human thinking, with 1000x runway ahead — alifcoder · 2026-09-22
• Qwen rumored to push 5-10T params with recursive self-improvement, plus Yunqi lineup — lxfater · 2026-09-22
• Qwen4 training now: Alibaba teases Max/Flash/Plus variants, 5-10T params for Qwen4.5-5 — ChrisGPT · 2026-09-22
• Alibaba Cloud Summit: Zhenwu V900 chip, Qwen4 teaser and a dozen AI launches — 智东西 · 2026-09-22
• Alibaba unveils new powerful chip and ambitious AI model plans — boppinmule · 2026-09-22
• Alibaba Roadmap: 20GW Global Compute by 2032, Zhenwu Chips 3x Faster — kevinsxu · 2026-09-22
• [source] Alibaba's Qwen 3.8 Max Runs on All-Domestic Zhenwu M890 Chips, Reports Claim — bookwormengr · 2026-09-22
• [source] Alibaba Unveils Zhenwu V900 AI Chip, Claims 3x Jump and China's Most Powerful — AIFlow_ML · 2026-09-23
• Alibaba's aggressive ASI roadmap: Qwen models to 5-10T params, 20GW by 2032 — bookwormengr · 2026-09-23
• Alibaba's Eddie Wu: Qwen sees RSI progress, plans 5-10T parameter model — teortaxesTex · 2026-09-23
• Alibaba roadmap: Qwen4 in training, Qwen4.5/Qwen5 up to 10T params, Qwen-Image-3.1 by year-end — julianharris · 2026-09-23
• Leaked Alibaba roadmap: Qwen 5 to scale to 5-10 trillion parameters — burny_tech · 2026-09-23
• Alibaba's roadmap: 5-10T parameter Qwen model, Zhenwu V900 chip, 20GW by 2032 — rohanpaul_ai · 2026-09-23
• Alibaba Unveils Zhenwu V900, Touted as China's Most Powerful AI Chip for 500K-GPU Clusters — emmanuelvivier · 2026-09-23
• Alibaba unveils Zhenwu V900 chip designed for 500,000-card AI clusters — emmanuelvivier · 2026-09-23

4 near-duplicate retellings: iamfakhrealam · robleclerc · emmanuelvivier · emmanuelvivier