Alibaba Releases Qwen3.8-Max, a 2.4 Trillion-Parameter Model Built for Multi-Day Autonomous Tasks

Alibaba's largest model yet targets days-long autonomy

Alibaba's Qwen team has released Qwen3.8-Max, a 2.4-trillion-parameter language model that activates 95 billion parameters per query. Built on the Qwen3.5 architecture, the model is designed to complete complex tasks independently over multiple days rather than respond to single prompts. It is available now through QwenCloud, with weights set to publish on Hugging Face and ModelScope next week — making it the first model in the Qwen-Max class to be open-weighted.

Alibaba first previewed the model in mid-July through its Token Plan, Qoder, and QoderWork at a discounted rate of ten percent of standard pricing. Exact per-token pricing has not been disclosed.

Autonomous coding and research reproduction

According to Alibaba, Qwen3.8-Max spent 16 days building a command-line tool called oh-my-cli entirely on its own, converting user requests into GitHub issues, writing code, running tests, and iterating. By July 30, 2026, the project had accumulated 265 commits, 127 pull requests, and 151 issues without human intervention, the company says.

In a second test, the model was given the paper "Unified Data Selection for LLM Reasoning" with no starter code. Over roughly five days and 125 hours of compute, it reproduced all six of the paper's main results, wrote 7,600 lines of code, ran 33 GPU training jobs, and — after testing 18 of its own ideas — claims to have beaten the paper's method on AIME24 by 2.7 points.

In a third case, the model competed in the WWW2025 Multimodal Dialogue Intent Recognition Challenge against 526 human teams on Alibaba's Tianchi platform. Within 24 hours it fine-tuned several models, including Qwen2.5-VL-7B, and pushed accuracy from 0.60 to 0.853 — outperforming 458 of 526 human teams, according to Alibaba.

Chip design and a simulated fiscal year

In a chip-design test, the model reduced a cryptographic circuit from 8,298 logic gates to 678 over about 500 iterations, shrinking physical chip area by 81 percent using the open-source layout tool OpenROAD, Alibaba says.

In E-Commerce-Bench, a year-long retail simulation built on anonymized Taobao and Tmall data, the model started with 100,000 yuan and ended with 416,252 yuan — a claimed 38 percent improvement over runner-up GLM 5.2 and 2.5 times its own predecessor, Qwen3.7-Max. The simulation included 152 hidden scam suppliers the model had to identify.

Benchmark claims and independent verification

Alibaba's published benchmark tables place Qwen3.8-Max near or above Claude Opus 4.8, Claude Fable 5, and GPT-5.6 Sol across several categories. The company reports a score of 93 on PaperBench, the highest in its comparison table, and 86.6 on TerminalBench 2.1, trailing GPT-5.6 Sol's 88.8. These are self-reported internal results; independent verification has not yet occurred.

Alibaba attributes the gains to an expanded reinforcement learning setup that trained on multi-day workflows, nested directory structures, and varied agent harnesses instead of single-task scenarios. The company's internal composite score across more than ten benchmarks reportedly rose from 0.474 to 0.725, peaking near 4,000 training environments before slightly declining.

The model supports both the OpenAI Chat Completions format and Anthropic's API protocol, allowing integration with Claude Code, Codex, Qoder CLI, Qwen Code, and OpenClaw. A reasoning_effort parameter lets users trade speed for depth across three settings.

Competitive context

Qwen3.8-Max's closest open-weight rival is Moonshot AI's Kimi K3, a 2.8-trillion-parameter multimodal MoE model with a one-million-token context window, released July 27. Independent testing reportedly showed Kimi K3 falling short of Western frontier models on cybersecurity and advanced math benchmarks.

What this means

Qwen3.8-Max represents a shift in how Chinese labs are positioning open-weight models — not as chatbots but as autonomous agents capable of running for days without supervision. The benchmark claims are notable, but they come exclusively from Alibaba's own testing, and multi-day agentic evaluations are especially hard to verify independently since they depend on tooling, environment setup, and reward design that outside researchers can't easily replicate. If the open weights match the claimed performance once released next week, Qwen3.8-Max would mark one of the largest openly available models to date and intensify competition with Kimi K3, DeepSeek, and GLM in the race for long-horizon agentic capability.