Alibaba Launches Qwen 3.8, a 2.4T-Parameter Open-Weight Model

The new multimodal model targets coding and complex productivity tasks, with open weights promised soon and a discounted preview already live.

Alibaba has released Qwen 3.8, an open-weight model with 2.4 trillion parameters that the company says trails only Fable 5 among frontier models [1].

For developers and businesses evaluating large model options, the release is notable on two fronts: scale and timing. Qwen 3.8 is the Qwen team’s first multimodal model exceeding 1 trillion parameters, capable of processing images, videos, and documents, according to developer Shuai Bai [1]. It also arrives as rival Moonshot AI’s Kimi K3 has been building commercial momentum [1].

What the model does

Alibaba says Qwen 3.8 should outperform its predecessor, Qwen 3.7-Max, particularly in coding and complex productivity tasks including full-stack development, data analysis, and office workflows [1]. No benchmark results have been published yet [1].

A preview is currently available through Alibaba’s Token Plan, Qoder, and QoderWork at 10 percent of the standard price [1]. Open weights are coming “soon,” the company said, though no specific date was given [1].

The competitive context

The release appears timed to put pressure on Moonshot AI, whose Kimi K3 model has attracted significant developer interest [1]. Moonshot has promised open weights of its own, but access currently runs through its chat app and application programming interface (API), effectively converting that interest into paying customers [1]. Qwen’s open-weight release could disrupt that strategy [1].

Moonshot reached $300 million in annual recurring revenue in June and plans to go public in as little as six months, according to Bloomberg’s reporting cited by The Decoder [1].


Sources

  1. The Decoder — Alibaba's Qwen takes on Kimi K3 with open-weight Qwen 3.8, says model is "second only to Fable 5"

This article was drafted with AI from the cited sources and checked against them before publication. Spot an error? Let us know.

A new framework from Alibaba researchers decomposes complex tasks iteratively so AI agents fetch only the tools they need, slashing context costs from ~884,000 tokens to ~1,160 per query.

Alibaba researchers have published SkillWeaver, a framework that routes AI agents to the right tools through iterative task decomposition rather than loading an entire tool library into context, reducing token consumption by over 99% compared to brute-force approaches [1].

For developers building enterprise agents that orchestrate dozens or hundreds of tools, the practical stakes are direct: fewer tokens mean lower API costs and faster responses, and the accuracy gains address a failure mode that currently makes large tool libraries nearly unusable [1].

The problem with loading every tool

Modern large language model (LLM) agents are increasingly integrated with massive tool ecosystems — including systems built on the Model Context Protocol (MCP), a standard for connecting agents to external services [1]. When an agent needs to handle a multi-step request like “Download the dataset, transform it, and create visual reports,” exposing the full tool library to the LLM to find the right tools is, as the researchers describe it, “highly inefficient” and “quickly overwhelms context limits” [1].

Most existing frameworks treat tool selection as a one-shot, single-skill problem, which breaks down when real business queries require sequencing multiple tools in a compatible chain [1].

How SkillWeaver works

SkillWeaver structures tool selection across three stages: Decompose, Retrieve, and Compose [1]. First, an LLM breaks a complex user query into atomic sub-tasks, each requiring exactly one tool [1]. An embedding model then searches the tool library to produce a shortlist of candidates for each sub-task [1]. Finally, a planner checks inter-tool compatibility — whether the output of one tool feeds cleanly into the next — and assembles the result as a Directed Acyclic Graph (DAG), allowing independent steps to run in parallel [1].

The central innovation is a feedback mechanism the team calls Skill-Aware Decomposition (SAD) [1]. Rather than decomposing a task once and searching, SAD has the LLM draft an initial plan, runs a preliminary search to surface loosely matching tools, and feeds those results back to the LLM so it can rewrite its decomposition using the actual vocabulary of the available tools [1]. The authors shared the prompt templates in their paper, and the technique can be reproduced with standard libraries such as LangChain, LlamaIndex, or plain Python [1].

Benchmark results

The team evaluated SkillWeaver on CompSkillBench, a custom benchmark of 300 multi-step queries run against a library of 2,209 real-world skills drawn from the public MCP ecosystem across 24 categories including cloud infrastructure, finance, and databases [1].

Without SAD, a 7-billion-parameter Qwen2.5-7B-Instruct model correctly predicted the number of required steps only 51.0% of the time; activating SAD pushed that to 67.7%, and using the larger Qwen-Max model brought it to 92% [1]. On hard tasks requiring four to five tools, SAD improved accuracy by 50% [1].

One counterintuitive finding: a 14-billion-parameter model performed worse than the 7B model in the unguided setup because it tended to over-decompose tasks into unnecessary micro-steps [1]. SAD corrected this by anchoring the larger model to the actual tool vocabulary, suggesting that tool alignment can matter more than raw model size [1].

The token savings were stark. The brute-force baseline — feeding all tool names into a large model’s prompt — consumed an estimated 884,000 tokens per query and still only retrieved the correct tool category 21.1% of the time [1]. SkillWeaver’s targeted approach brought that figure down to roughly 1,160 tokens per query, a 99.9% reduction, while substantially outperforming on accuracy [1]. A ReAct-style agent baseline achieved 0% decomposition accuracy, collapsing multi-step plans into isolated actions [1].

On retrieval infrastructure, the researchers used the open-source all-MiniLM-L6-v2 embedding model with a FAISS index [1]. Indexing all 2,209 skills took 15 seconds, and retrieval adds under 15 milliseconds of latency per query [1]. Swapping in a stronger encoder (BGE-base-en-v1.5) improved accuracy without any fine-tuning [1].

Limitations and what’s not included

The framework currently has no error-recovery mechanism [1]. If a tool call fails mid-chain, the entire execution breaks, and the authors acknowledge that practitioners will need to build their own retry and fallback logic on top of the compose stage [1].

The source code has not yet been released, though the researchers note the system was built entirely from off-the-shelf components and the SAD prompt templates are available in the paper [1].


Sources

  1. VentureBeat — New Alibaba AI framework skips loading every tool, cutting agent token use 99%

This article was drafted with AI from the cited sources and checked against them before publication. Spot an error? Let us know.

Anthropic says Alibaba-linked operators ran 28.8 million exchanges through nearly 25,000 fraudulent accounts to extract Claude's capabilities, and is urging Congress to act.

Anthropic has accused Chinese technology giant Alibaba of orchestrating the largest campaign ever recorded to illicitly copy, or “distill,” the capabilities of its Claude large language model (LLM), according to a letter the AI company sent to two U.S. senators on June 10. [1]

The letter, addressed to Senators Tim Scott (R-SC) and Elizabeth Warren (D-Mass.) and obtained by Ars Technica, was sent one day before a Senate committee hearing on “AI and the American Dream.” [1] In it, Anthropic described “new, confidential evidence of the largest campaign to illicitly extract Claude’s capabilities we have ever measured.” [1]

According to Anthropic, the attacks took place between April 22 and June 5, during which “operators affiliated with Alibaba and Alibaba Qwen, Alibaba’s AI lab” generated “more than 28.8 million exchanges with Claude through almost 25,000 fraudulent accounts.” [1] The campaign specifically targeted capabilities including agentic reasoning, software engineering, and long-horizon tasks, Anthropic said. [1]

Anthropic alleged that Alibaba evaded detection by “using obfuscation techniques and proxy networks,” and warned that a “growing circumvention economy” is emerging to support future distillation attacks. [1] Distillation attacks involve using a target model’s outputs to train a competing model, effectively transferring capabilities without incurring the original research and development costs. [1]

The alleged campaign unfolded after President Donald Trump publicly warned against such activity. In April, Trump accused China of “industrial-scale” AI theft following earlier Anthropic allegations that Chinese firms DeepSeek, Moonshot, and MiniMax had used similar tactics to generate over 16 million exchanges with Claude through approximately 24,000 fraudulent accounts. [1] OpenAI and Google have also published findings on comparable attacks against their own models, Anthropic noted. [1]

Anthropic pointed out that Alibaba is listed on the New York Stock Exchange and maintains U.S. business operations, yet the alleged activity continued after Trump’s memo warned that cloning attempts were “unacceptable.” [1] Anthropic characterized Alibaba as “brazenly” racing to produce a copycat model despite its exposure to U.S. investors and regulators. [1]

Alibaba has not been passive in its own disputes with the U.S. government. In a lawsuit filed this week, Alibaba accused the Trump administration of wrongly blacklisting the company by linking it to the Chinese military, Reuters reported. [1] Alibaba stated that “its products and services are built for retail, logistics, and enterprise information technology—not weapons, defense, or intelligence.” [1] Alibaba’s stock fell 3 percent after Anthropic’s accusations became public, according to Yahoo Finance. [1]

Anthropic’s letter to Congress outlined three legislative recommendations. First, antitrust laws should be updated to allow AI companies to share intelligence about Chinese tactics. [1] Second, expanded export controls on advanced chips could limit China’s ability to train on U.S. model outputs, potentially making distillation attacks less worthwhile. [1] Third, Congress should pass laws penalizing Chinese labs’ conduct, potentially restricting their access to U.S. models, advanced chips, or data centers outside China. [1]

Anthropic declined to say whether the Alibaba campaign was significant enough to meaningfully accelerate China’s AI capabilities, or to detail specific countermeasures taken. [1] In a statement to Ars Technica, the company said: “We believe combating the threat of illicit distillation requires coordinated action between government and industry, and we will continue working with Congress and the Administration to maintain American AI leadership.” [1]

The strategic stakes Anthropic describes extend beyond commercial competition. The company warned that if China closes the capability gap, it could deploy “advanced cyber capabilities” against U.S. government and corporate targets and exploit vulnerabilities faster than previously possible. [1] Anthropic also cautioned that China could release advanced models with weak safety guardrails that are easily manipulated, enabling U.S. adversaries to use them for harmful purposes. [1]

Independent corroboration of China’s urgency came from a cybersecurity conference in Beijing, where 360 Security Technology founder Zhou Hongyi publicly described Anthropic’s Mythos model — which has been restricted from foreign markets — as a “cyber nuclear weapon” and a “terrifying change” in the ability to find security vulnerabilities. [1] Zhou said China’s exclusion from Project Glasswing, a U.S. initiative that gave more than 40 American organizations access to Mythos Preview for cyber defense purposes, left China at a strategic disadvantage. [1] “This means US organizations can use Mythos to scan your vulnerabilities, but you don’t even have the qualification to catch a glimpse of Mythos,” Zhou said. [1] He argued China must develop its own equivalent capability to achieve a form of mutually assured deterrence. [1]

Alibaba’s Qwen family of AI models has been downloaded more than 700 million times, and China’s state-run People’s Daily has described it as “the most popular open-source AI system worldwide.” [1]


Sources

  1. Ars Technica — Anthropic says Alibaba must be punished for largest Claude cloning attack

This article was drafted with AI from the cited sources and checked against them before publication. Spot an error? Let us know.