Alibaba’s Qwen team announced on July 19, 2026 that Qwen3.8 is launching with a claimed 2.4 trillion parameters, with a preview already available for purchase and open weights promised for a later date. The announcement came during the World AI Conference (WAIC) in Shanghai.
Alibaba described Qwen3.8 as “one of the most powerful models available today, comparable to leading frontier AI models, second only to Fable 5,” referring to Anthropic’s flagship. The preview build, labeled Qwen3.8-Max-Preview, is live on Alibaba’s Token Plan, Qoder, and QoderWork platforms. No benchmark table, model card, license, or active-parameter count has been published.
The timing is significant. Qwen3.8 arrived just two days after Moonshot AI released Kimi K3, a 2.8 trillion-parameter open-weight model whose launch helped trigger a broad semiconductor selloff that erased roughly $3.3 trillion in global chip market value within weeks. It also followed the July 15 regulatory approval for Apple Intelligence powered by Qwen in China, meaning attention on the Qwen brand was already elevated.
What Is Qwen3.8-Max
Qwen3.8-Max is a multimodal large language model from Alibaba Cloud’s Qwen team, described as a sparse mixture-of-experts (MoE) architecture with 2.4 trillion total parameters. The total parameter count is Alibaba’s own figure and has not been independently confirmed.
Qwen developer Shuai Bai described it as the team’s first multimodal model above 1 trillion parameters. The model follows Qwen3-Max (estimated 1T, September 2025), Qwen3.6-Max-Preview (April 2026, closed), and Qwen3.7-Max (May 2026, closed but with published benchmarks). Qwen3.8-Max-Preview is the first in the series to ship without any benchmark table.
One detail Alibaba has not disclosed is the number of active parameters per token. In a sparse MoE model, only a fraction of total parameters activate during inference. Alibaba’s own Qwen3-235B, for example, carries 235 billion total parameters but activates only 22 billion per token. Without this number for Qwen3.8, the 2.4T figure describes the model’s size on disk, not its computational cost per query.
Don’t Confuse Qwen3.8 With Qwen3-8B
Search engines are currently conflating two very different models, so this distinction matters.
Qwen3.8-Max is Alibaba’s 2.4 trillion-parameter flagship announced on July 19, 2026. It is a sparse MoE model available only through Alibaba’s cloud platforms. Open weights have been promised but not delivered.
Qwen3-8B is a dense, 8.2 billion-parameter open-weight model released in April 2025 under the Apache 2.0 license. It has been available on Hugging Face for over a year.
The two models are roughly 300 times apart in parameter count and serve entirely different use cases. Qwen3-8B runs on consumer hardware. Qwen3.8-Max, at 4-bit quantization, would require approximately 1.2 terabytes of memory for weights alone. A single Nvidia H200 carries 141 GB. Running Qwen3.8 locally is not a realistic option for independent developers.
How to Access Qwen3.8 and What It Costs
Qwen3.8-Max-Preview is available through three platforms, all operated by Alibaba.
- Alibaba Cloud Token Plan offers three subscription tiers. The Lite plan starts at 39 CNY per month (approximately $6) with 2,500 credits per seven-day cycle. The Standard plan costs 139 CNY. The Pro plan costs 499 CNY and includes 40,000 credits per cycle with support for six to eight concurrent agents.
- Qoder, Alibaba’s AI coding platform, includes Qwen3.8-Max-Preview access at reduced credit consumption during off-peak hours.
- QoderWork, Alibaba’s task automation platform, also supports the preview.
Standalone per-token API pricing has not been published separately from the Token Plan bundles. Preview pricing runs at 10% of standard rates, according to MarkTechPost. The Token Plan also includes access to Qwen3.7-Max, DeepSeek V4 Pro, and GLM-5.2.
Regarding open weights and Qwen3.8 release date, Alibaba’s announcement stated that open weights are coming “soon.” No specific date, license terms, or Hugging Face repository has appeared. Alibaba’s last two Max-tier flagships, Qwen3.6-Max and Qwen3.7-Max, both shipped as closed models despite the company’s historical practice of open-weighting its top releases.
The 2.4 Trillion-Parameter Reality Check
The 2.4T number deserves careful context. Total parameters in a sparse MoE model do not translate directly into performance or compute cost. They describe how large the model is when stored, not how much processing happens when it generates a response.
For comparison, DeepSeek V4 Pro carries 1.6 trillion total parameters but activates roughly 49 billion per token, around 3% of the network. If Qwen3.8 follows a similar ratio, its effective compute per token could be a small fraction of what the headline number suggests.
Alibaba’s claim that Qwen3.8 is “second only to Fable 5” is based on internal evaluation. No independent benchmark results have been published. The most recent Qwen model with a full public benchmark table remains Qwen3.7-Max from May 2026. Alibaba’s documentation also notes that Qwen3.8-Max-Preview is a “continuously evolving” build, meaning model behavior may change during the preview window.
Qwen3.8 in the Trillion-Parameter Race
Qwen3.8’s announcement adds another entry to a rapid sequence of trillion-parameter Chinese AI models released in 2026.
- DeepSeek V4 Pro (April 2026) shipped with 1.6 trillion total parameters, approximately 49 billion active, and published benchmarks
- Kimi K3 (July 17, 2026) launched with 2.8 trillion parameters and open weights, with benchmark results placing it near Fable 5 and GPT-5.6
- Qwen3.8-Max-Preview (July 19, 2026) claims 2.4 trillion parameters with no benchmarks and no open weights yet
Kimi K3’s release carried immediate market consequences. The Philadelphia Semiconductor Index fell more than 20% from its June peak, entering bear market territory. Bloomberg reported the selloff erased approximately $3.3 trillion in chip market value since June 22. Moonshot AI, the company behind Kimi K3, is backed by Alibaba, creating an unusual dynamic where both the 2.8T and 2.4T entries in this race share a common investor.
The Security Question
One angle that most coverage has not examined closely involves data handling. Qwen3.8-Max-Preview currently runs entirely on Alibaba’s infrastructure. Under China’s National Intelligence Law, Chinese companies are required to cooperate with government data requests, a provision that has drawn scrutiny from Western security researchers evaluating Alibaba’s AI coding tools.
This concern is not theoretical. Earlier in July, Alibaba banned its employees from using Anthropic’s Claude Code and directed them to use Qoder instead, after allegations that Claude Code contained hidden detection features targeting Chinese users. Anthropic, in turn, had accused operators linked to Alibaba’s Qwen lab of running a large-scale distillation campaign against Claude. For developers considering Qwen3.8-Max-Preview for production workloads, this geopolitical context is a practical consideration alongside performance and pricing.
What to Watch Next
Three variables will determine whether Qwen3.8 lives up to its positioning.
Open weights. If Alibaba delivers downloadable weights, Qwen3.8 would become the largest open-weight model ever released. If the weights do not materialize, as happened with Qwen3.6-Max and Qwen3.7-Max, the promise joins a growing list of unfulfilled commitments in 2026.
Independent benchmarks. Until third-party evaluations appear, comparisons with Fable 5 or Kimi K3 remain speculative. Alibaba shipped a buyable preview before it shipped a single benchmark result.
Active parameter disclosure. The number that determines real serving cost and practical viability for developers remains unpublished. This single figure will shape whether Qwen3.8 represents a genuine compute advance or a parameter-count headline.
This article was last updated on July 20, 2026. Information may change as Alibaba publishes additional technical documentation.
FAQs
What is Qwen3.8-Max-Preview?
Qwen3.8-Max-Preview is a multimodal large language model from Alibaba with a claimed 2.4 trillion parameters, announced on July 19, 2026. It is available through Alibaba’s Token Plan subscription, Qoder, and QoderWork, but no independent benchmarks or model card have been published.
When will Qwen3.8 open weights be released?
Alibaba stated that open weights are coming “soon” but has not provided a specific date, license type, or Hugging Face repository. The company’s two previous Max-tier flagships, Qwen3.6-Max and Qwen3.7-Max, both launched as closed models despite similar early promises.
Is Qwen3.8 the same as Qwen3-8B?
No. Qwen3.8-Max is a 2.4 trillion-parameter flagship model announced in July 2026. Qwen3-8B is a separate 8.2 billion-parameter open-weight model released in April 2025. The two are roughly 300 times apart in parameter count and serve different use cases.
How does Qwen3.8 compare to Kimi K3?
Both are trillion-parameter models from Chinese AI labs, but they differ in key ways. Kimi K3 has 2.8 trillion parameters, published benchmarks, and released open weights. Qwen3.8 claims 2.4 trillion parameters but has published no benchmarks and no weights. Direct performance comparison is not possible until independent testing data becomes available.
The post What Is Qwen3.8? Alibaba’s 2.4T Model Explained appeared first on Memeburn.