llms.blog
AI Models

Alibaba launches Qwen 3.8-Max, benchmarks fall short of claims

Alibaba officially launched Qwen 3.8-Max on August 3, calling it the most powerful model in the Qwen family to date. The 2.4 trillion-parameter model supports a one-million-token context window and is multimodal, capable of processing text, images, video, and documents in a single prompt. The company priced Qwen 3.8-Max at $2 per million input tokens and $6 per million output tokens, with implicit caching at $0.25 per million tokens. That positions it well below the pricing of frontier models f

axc
axc
2 min
Alibaba launches Qwen 3.8-Max, benchmarks fall short of claims

Alibaba officially launched Qwen 3.8-Max on August 3, calling it the most powerful model in the Qwen family to date. The 2.4 trillion-parameter model supports a one-million-token context window and is multimodal, capable of processing text, images, video, and documents in a single prompt.

The company priced Qwen 3.8-Max at $2 per million input tokens and $6 per million output tokens, with implicit caching at $0.25 per million tokens. That positions it well below the pricing of frontier models from Anthropic and OpenAI. Open weights are promised for next week, alongside a smaller Qwen 3.8-27B variant.

Architecture and capabilities

Qwen 3.8-Max uses a mixture-of-experts design. Although the model has 2.4 trillion total parameters, it activates approximately 95 billion per query, which keeps inference costs down while retaining a large knowledge base. Alibaba says the model is designed for complex reasoning, autonomous coding, multimodal understanding, and long-running agent tasks.

In an internal demonstration, Alibaba reported that Qwen 3.8-Max completed a software engineering project over 16 days with minimal human supervision, handling planning, coding, debugging, and optimization. The company also said the model can reconstruct applications from screenshots, convert 2D floor plans into 3D visualizations, and build searchable knowledge bases from mixed media.

Benchmark reality check

Alibaba claimed the model is "second only to Fable 5" and shared benchmark results showing strong performance in coding, software engineering, research, long-video understanding, and computer-use tasks. The company said Qwen 3.8-Max is the highest-ranked Chinese text model on Arena.AI and ranks second globally for visual understanding, behind Anthropic's Claude Fable 5.

However, Nikkei Asia reported on August 3 that independent benchmark testing puts Qwen 3.8-Max behind several domestic and foreign rivals, falling short of Alibaba's "second only to Fable 5" claim. The benchmark results shared by Alibaba have not been independently verified.

Competitive context

The launch comes days after Moonshot AI released Kimi K3, a 2.8 trillion-parameter open-weight model. Qwen 3.8-Max is priced significantly below Kimi K3, continuing a price war among Chinese AI labs. DeepSeek also recently moved V4 into general availability with dynamic off-peak pricing of $0.87 per million tokens, compared to $50 per million for Anthropic's Fable 5.

The sequence of releases from Moonshot, DeepSeek, and Alibaba underscores China's sustained push to compete at the frontier of AI model development, particularly on cost and open-weight availability.

Sources

Alibaba launches its most powerful model Qwen 3.8-Max - India Today, August 3, 2026: https://www.indiatoday.in/technology/news/story/alibaba-launches-its-most-powerful-model-qwen-38-max-claims-it-can-match-performance-of-chatgpt-and-claude-2962176-2026-08-03

Alibaba's new Qwen AI model falls short of 'second only to Fable 5' claim - Nikkei Asia, August 3, 2026: https://asia.nikkei.com/business/technology/artificial-intelligence/alibaba-s-new-qwen-ai-model-falls-short-of-second-only-to-fable-5-claim

Alibaba's Qwen takes on Kimi K3 with open-weight Qwen 3.8 - The Decoder, July 19, 2026: https://the-decoder.com/alibabas-qwen-takes-on-kimi-k3-with-open-weight-qwen-3-8-says-model-is-second-only-to-fable-5

DeepSeek and Alibaba launch fresh assaults on frontier AI - Cybernews, July 20, 2026: https://cybernews.com/ai-news/deepseek-and-alibaba-launch-fresh-assaults-on-frontier-ai

AI ModelsAlibabaChina
ShareX (Twitter)LinkedIn

Written by

axc

More to read

  • Zero-Downtime Model Updates in Production LLM Serving: In-Place Weight Transfer, CUDA IPC vs. NCCL Syncing, and Traffic Draining Architectures

    Zero-Downtime Model Updates in Production LLM Serving: In-Place Weight Transfer, CUDA IPC vs. NCCL Syncing, and Traffic Draining Architectures In high-throughput LLM serving infrastructure, updating model checkpoints presents a severe operational dilemma. Traditional microservice deployment patterns such as blue-green deployments or rolling pod restarts fail to scale economically when applied to multi-GPU LLM clusters. Tearing down a serving instance to load a newly fine-tuned checkpoint or pol

    1 min

  • The Curse of Multilinguality in Large Language Models: Capacity Dilution, Tokenizer Fertility, and Representation Interference

    The Curse of Multilinguality in Large Language Models: Capacity Dilution, Tokenizer Fertility, and Representation Interference Training a single transformer foundation model to process dozens or hundreds of languages is one of the central goals of modern natural language processing. In theory, massive multilingual pre-training unlocks positive cross-lingual transfer: low-resource languages gain syntactic, factual, and reasoning capabilities from the rich supervision available in high-resource l

    1 min

  • LLM Load Shedding and Overload Control in Production: Adaptive Admission, Preemption Economics, and Graceful Degradation

    LLM Load Shedding and Overload Control in Production: Adaptive Admission, Preemption Economics, and Graceful Degradation Standard web services rely on well-established overload protection patterns: reverse proxies monitor CPU utilization, memory thresholds, or static queue depths and reject excess HTTP requests with 429 Too Many Requests or 503 Service Unavailable status codes. When applied naively to Large Language Model (LLM) serving infrastructure, these conventional heuristics fail catastro

    1 min