Alibaba's Qwen 4 Is Coming — Here's What the Architecture Preview Actually Tells Us

# Alibaba's Qwen 4 Is Coming — Here's What the Architecture Preview Actually Tells Us

*Wei Lian — October 20, 2026*

Alibaba has confirmed only a narrow set of facts about Qwen 4: the model is in training, it belongs to a roadmap extending through Qwen 4.5 and Qwen 5, and that roadmap contemplates models in the 5–10-trillion-parameter range. Everything more specific — from the release day to the production model's capabilities and deployment economics — remains unconfirmed.

The best technical evidence is therefore not a prediction-market price or a parameter target. It is Qwen3.8-Flash-Next, an open-weight model that the Qwen team explicitly describes as an early architectural preview of Qwen 4. That preview points toward an unusually sparse design: limited accelerator-active computation, substantial capacity held in host memory, and long-context mechanisms intended to reduce attention costs. Alibaba Cloud is simultaneously preparing the Zhenwu V900 accelerator and much larger clusters, but the chip's headline performance remains a company claim unsupported by published FLOPS, process-node, or power figures.

What Alibaba Has Actually Confirmed

At its Apsara Conference on September 22, 2026↗, Alibaba said Qwen 4 was currently in training. The company also placed Qwen 4.5 and Qwen 5 on the roadmap and said future series were projected to scale to between 5 trillion and 10 trillion parameters. Alibaba did not provide a calendar release date or publish a Qwen 4 model card, technical report, API specification, or weights.

Secondary reporting attributes the conference statement to Qwen project lead Liu Dayiheng and says the model would arrive "very soon." The official Apsara release, however, does not identify the speaker responsible for the Qwen 4 statement. The more defensible formulation is that Alibaba announced the training status at Apsara; the attribution to Liu comes from secondary market-tracking material rather than the official release.

The same distinction applies to the proposed product lineup. Tracking sources report four anticipated tiers:

• Qwen 4 Max — the flagship model, targeting high-level reasoning and agentic tasks
• Qwen 4 Plus — a balanced, multimodal tier for general enterprise use
• Qwen 4 Flash — optimized for low-latency, high-volume, cost-sensitive applications
• Qwen 4 27B — an open-weights model specifically designed for local deployment on consumer hardware

These tiers should be treated as a reported lineup, not confirmed launch inventory. Alibaba's official Apsara account does not provide them as a formal product specification.

The decisive questions begin only after shipment. Qwen 4 should be judged by independently evaluated capability, sustained serving speed, real context performance, hardware and geographic availability, open-weight terms where applicable, and total economics.

Flash-Next Is the Meaningful Preview

On August 26, the Qwen team released Qwen3.8-Flash-Next↗ and called it "an early preview of the architecture used in Qwen4." The team said it was exposing the changes before building the full Qwen 4 family on top of them. This language makes Flash-Next the strongest public clue to architectural direction, while stopping short of proving the final Qwen 4 configuration.

The Hugging Face model card↗ and the technical report on arXiv↗ describe a model that combines a 125-billion-parameter sparse mixture-of-experts backbone with only 6 billion parameters activated per token. It adds 51 billion parameters of n-gram embeddings held off the accelerator in host memory, plus a 4-billion-parameter multi-token-prediction component. The model supports 262,144 tokens natively and can be extended to 1 million tokens using YaRN scaling.

The Four Design Choices That Define the Preview

Four architectural decisions stand out as strategically significant:

• Sparse MoE backbone — 125B total parameters, 6B active per token, dramatically reducing per-token compute relative to a dense model of equivalent total capacity
• N-gram embedding table — 51B parameters held in host RAM and prefetched on demand, adding lookup-based capacity without increasing accelerator-memory pressure
• Hybrid attention (GDN + QSA) — alternating Gated DeltaNet recurrent layers with Qwen Sparse Attention retrieval layers, targeting efficient long-context handling
• Gated Residual pathways with Muon optimizer — widening the residual stream into four branches with data-dependent gates, improving representational capacity and training stability

The attention stack alternates three Gated DeltaNet layers with one Qwen Sparse Attention layer in each group of four. Gated DeltaNet compresses prior context into a fixed-size recurrent state, while QSA uses a lightweight indexer to score and select context at micro-block granularity. According to Intuition Labs' architecture analysis↗, Qwen reports that at a 1-million-token context, its QSA kernel was 7.6 times faster in prefill and 4.9 times faster in decode than dense attention. Those are company-reported kernel results, not independent demonstrations of complete serving economics.

The N-Gram Table: Capacity Without Compute

The n-gram layer is the most strategically distinctive element. Its 51 billion parameters are held off the accelerator and prefetched from host memory, allowing the model to add lookup-based capacity with little incremental per-token computation. The MindStudio overview↗ notes that the vLLM deployment guide↗ explicitly recommends enabling CPU offload for the n-gram table on GPUs with limited VRAM, such as 80GB H100s, to prevent out-of-memory errors.

The team's own experiments contain an important qualification: expanding the n-gram vocabulary reduced loss monotonically, but downstream accuracy eventually saturated. More offloaded parameters therefore do not automatically produce proportionate capability gains.

The n-gram table represents a bet that linguistic pattern memory can be separated from reasoning compute — a design philosophy that could reshape how future Chinese frontier models are deployed on heterogeneous hardware.

Zhenwu V900 Supplies the Hardware Thesis

Alibaba's Qwen roadmap is paired with Zhenwu V900, an accelerator designed by its T-Head unit for training and inference. As Tom's Hardware reported↗, Alibaba specifies 216GB of memory, 1,200GB/s of inter-chip bandwidth, and native FP8 and FP4 support. An upgraded supernode architecture can support clusters comprising as many as 500,000 cards, with mass production and commercial release scheduled for Q1 2027 according to TrendForce↗.

Alibaba CEO Eddie Wu called the V900 China's most powerful AI chip and said it delivered three times the performance of the preceding M890. These are Alibaba claims. Tom's Hardware noted that the disclosed specifications omit FLOPS, manufacturing process, foundry, and power consumption, preventing a complete performance or efficiency comparison.

The V900 matters less as proof of current leadership than as evidence of vertical integration. Alibaba is aligning model sparsity, low-precision computation, interconnect bandwidth, and cluster scale under one roadmap. The inference is that future Qwen systems may be co-designed around domestic hardware rather than merely ported to it — a strategic response to US export controls that restrict access to advanced NVIDIA chips.

V900 Specifications at a Glance

• Memory: 216GB per card — disclosed specification
• Inter-chip bandwidth: 1,200GB/s — disclosed specification
• Numeric formats: FP8 and FP4 native support
• Cluster scale: Up to 500,000 cards via T-Head ICN Switch
• Performance claim: 3× the M890 — Alibaba-reported, unverified by third parties
• Commercial timeline: Q1 2027 mass production; J900 successor planned for Q3 2027

Competitive Landscape: Where Qwen Stands Today

The available Artificial Analysis Intelligence Index↗ snapshot places the current Chinese model field in a tight cluster, with Xiaomi's MiMo-V2.6-Pro leading at a score of 46, followed by Qwen3.8-Max and Zhipu AI's GLM-5.3 at 45, and StepFun's Step 5 Preview at 44. These are current-model comparisons, not evidence of Qwen 4 performance.

The gap to the Western frontier remains meaningful. Anthropic's Claude Opus 5.5 scores 57.6 and OpenAI's GPT-6.1 Sol scores 52 on the same index — roughly an 11-to-12-point lead over the best Chinese models. The Center for AI Standards and Innovation estimates that top Chinese models lag the US frontier by approximately eight months in capability terms.

Alibaba separately says automated recursive self-improvement raised Qwen3.8-Max's Artificial Analysis score from 40 to 45 across 33 iterative cycles in one month. That is a company-reported development result. It should not be converted into a claim that Qwen 4 will inherit the same score trajectory or outperform listed competitors.

Capital and Policy Constraints

Alibaba's infrastructure commitment supplies financial context. On February 24, 2025, the company committed at least RMB380 billion over three years↗ to AI and cloud infrastructure — approximately $52–53 billion↗ at contemporary exchange rates. This is a forward investment commitment, not proof that the full amount has already been spent.

The policy backdrop is directly relevant. US restrictions on advanced AI chips and chipmaking equipment constrain Chinese access and increase the strategic importance of domestic accelerators such as the V900. The recently proposed Chinese export controls on advanced AI models and weights — still in draft form as of this writing — add a further layer of uncertainty for developers planning to build on open-weight Chinese models.

What Developers Should Watch For

The YottaLabs Qwen 4 tracker↗ reported that as of October 12, prediction markets assigned approximately a 75% probability to a Qwen 4 launch before November 1, 2026. That number is a time-stamped indicator of trader sentiment, not an Alibaba forecast.

For developers and enterprises evaluating whether to build on Qwen 4, the practical checklist is straightforward:

• Open-weight availability: Will the 27B tier ship under Apache 2.0, MIT, or a more restrictive custom license? Alibaba's Qwen3.8-Max used a custom license with revenue thresholds; the 27B model may follow a more permissive path.
• Hardware requirements: The YottaLabs 27B hardware guide↗ estimates that a dense 27B model would fit on a 24GB GPU at 4-bit quantization, a 48GB GPU at FP8, and an 80GB GPU at full BF16 — but if the 27B adopts the Flash-Next MoE architecture, memory planning becomes more complex due to the n-gram table.
• API pricing: Qwen3.8-Max's current API pricing and the Flash-Next preview's inference economics will set expectations; Qwen 4 Flash will likely undercut both.
• Benchmark independence: Vendor-reported scores on SWE-bench Pro (62.5%), CoWorkBench (73.9%), and AndroidWorld (84.5%) for the Flash-Next preview are a starting point, not a final verdict. Independent evaluations on Artificial Analysis and LMSYS will matter more.

The Pre-Launch Test

The evidence supports a coherent direction: Qwen 4 is in training; Alibaba is exploring much larger future models; Flash-Next previews an architecture built around sparse activation, offloaded n-gram capacity, and cheaper long-context retrieval; and V900 represents an attempt to control more of the compute stack. It does not support a confirmed Qwen 4 release date, final tier list, benchmark score, parameter count, or cost profile.

Roadmap scale can signal ambition, and traders' odds can summarize sentiment. Neither substitutes for a model that users can test, deploy, and price. The Qwen 4 story begins in earnest the moment Alibaba publishes weights, an API, and a technical report — and independent evaluators have had time to run their own benchmarks.