Alibaba Qwen3.8-Max: 2.4T-Parameter Model Targets OpenAI, Anthropic

Alibaba's biggest bet yet on open frontier AI

On August 3, 2026, Alibaba made Qwen3.8-Max — a 2.4-trillion-parameter flagship model and the largest system it has ever shipped — generally available through QwenCloud and Alibaba Cloud's Model Studio, with open weights promised for the following week. It is the first time Alibaba has put a Max-class model into open-weight territory, reversing the company's recent practice of keeping flagship releases proprietary16. Investors read it as a signal: Alibaba's Hong Kong-listed shares closed up 7 percent on the day, and its New York premarket trading rose 4.5 percent15.

The model had been previewed on July 19 at the World AI Conference in Shanghai under the name Qwen3.8-Max-Preview, landing just days after Moonshot AI announced its own 2.8-trillion-parameter Kimi K3 — timing that many in the community read as a direct response to that rival's open-weight positioning71.

Scale without the serving bill

Under the hood, Qwen3.8-Max is a sparse mixture-of-experts model built on the Qwen 3.5 architecture, paired with a hybrid attention mechanism. The headline 2.4-trillion-parameter count is somewhat misleading as a cost indicator: only about 95 billion parameters — roughly 4 percent of the total — activate during inference, which is what makes a model of this scale servable at all, and puts it in a lighter serving class than Kimi K3's 104 billion active parameters128. The context window stretches to 1 million tokens — about 750,000 words per query — with a detailed spec breakdown putting maximum input near 991,000 tokens, output at 131,000, and a 262,000-token reasoning ceiling1614.

The model is natively multimodal, accepting text, image and video input, and Alibaba has leaned hard into document and video ingestion: the company says it can digest financial reports exceeding 200 pages and footage running past 100 hours, turning that material into searchable knowledge bases56. It can also reconstruct applications from a single UI screenshot, build interactive games from natural language, and convert 2D floor plans into 3D visualizations18.

The autonomous-coding pitch

What Alibaba clearly wants the market to remember, though, is the long-horizon agentic story. In internal testing, the company says the model ran fully autonomously for more than 16 days on a self-evolving software project — accumulating 265 commits, 127 pull requests and 151 issues in a repository with no human commits — building a self-improving agent framework it calls "oh-my-cli," which has since been open-sourced on GitHub58. In another test, the model entered a 526-team machine learning competition under a 24-hour limit and finished ahead of roughly 87 percent of the human field510.

Alongside the model, Alibaba launched QwenWork, an all-in-one workplace AI agent platform that entered public beta the same day via web and desktop apps, combining Alibaba's existing agent tools and targeting Tencent's WorkBuddy, Moonshot's Kimi Work, and overseas rivals including Claude Cowork and ChatGPT Work, with DingTalk integration to follow15.

A benchmark table that cuts both ways

Alibaba framed the launch with the claim that Qwen3.8-Max is "second only" to Anthropic's Fable 5, and the release finally shipped with a full benchmark table — something the July preview lacked17. On the crowdsourced Arena.AI platform, the model immediately became the highest-ranking Chinese model for text tasks, ranking fifth on Text Arena and second on Vision Arena, trailing only select Anthropic offerings318.

The detailed numbers tell a more nuanced story than the marketing. Qwen3.8-Max scores 86.6 on Terminal-Bench 2.1, ahead of Claude Opus 4.8 and Claude Fable 5 at 84.6, but behind GPT-5.6 Sol at 88.8. It posts strong multimodal and instruction-following results — 92.6 on GPQA Diamond, 82.8 on IFBench, 86.1 on OSWorld-Verified — but it trails Fable 5 badly on the hardest software-engineering test, SWE-bench Pro, at 67.7 versus 80.0, and sits last among the four compared flagships on Humanity's Last Exam at 43.669.

That pattern — winning on agentic terminal work, instruction following and multimodal reasoning, losing on broad knowledge and the deepest engineering — is consistent across the published table, and it matters for buyers deciding between Alibaba and the closed US labs. Crucially, every number is vendor-run; no independent evaluator had scored the model as of late August, and most coding rows were produced by running Qwen3.8-Max inside Anthropic's own Claude Code harness via its Anthropic-compatible API — a fair apples-to-apples setup, but one where harness choice can swing results915.

Pricing as the real headline

For many buyers, the pricing is doing more work than the benchmarks. Qwen3.8-Max launched at $2 per million input tokens and $6 per million output tokens, with cached input at $0.25 — less than a third of Claude Opus 5's combined rate and under a quarter of GPT-5.6 Sol's standard-mode pricing, while also undercutting Kimi K3's $3/$6/$15 schedule29. For enterprises feeling locked into a single closed vendor, a near-frontier model that anyone can download or call for a fraction of the price changes the procurement calculus even for teams that never switch14.

The API supports both OpenAI-compatible endpoints and an Anthropic-compatible interface, meaning Claude Code, Codex and similar tools can point at Qwen3.8-Max with minimal configuration changes46.

Where the open-weight release actually landed

The open-weight drop came in two pieces in mid-August. Qwen3.8-27B, aimed at ordinary on-premise GPU hardware, hit Hugging Face under Apache 2.0, while the Max-class 2.4T checkpoint — Qwen3.8-2.4T-A95B — went up under a custom license, with one significant caveat: the open release is text-only, lacking the vision capabilities and the native 1M context of the hosted flagship2. Alibaba had not disclosed license terms at launch, and coverage diverged on how open the "open" release would ultimately be46.

The move nonetheless marks Alibaba's return to open-sourcing its top-tier models after keeping several 2026 flagships proprietary, and it comes amid a broader shift in Chinese AI competition toward coding agents, multimodal systems and multi-step workflow execution — with DeepSeek upgrading its V4-Flash API for agent capabilities and Moonshot's Kimi K3 as the direct domestic rival13.

The read: pressure, not parity

The strongest interpretation of the evidence is this: Alibaba has not overtaken OpenAI or Anthropic, and its own benchmark table says so — Fable 5 and GPT-5.6 Sol still lead on the hardest reasoning and engineering work. What Alibaba has done is collapse the price-and-openness gap to the point where "frontier-adjacent" is no longer a compromise. A downloadable, self-hostable 2.4T-parameter model within two points of closed frontier systems on terminal-based coding — updated within a month by the Qwen3.8-Max-0902 snapshot that topped Code Arena WebDev at 1,691 points, edging Claude Opus 5 Max by 3 — is a genuine competitive problem for US labs414.

The compression of release cycles is the real story: Qwen 3.7-Max shipped in May, the 3.8 preview in July, general availability in August, open weights the same month, and a post-training upgrade in September, while OpenAI, Anthropic and Google all shipped updates within days of each other in early September214. Whether the benchmark claims hold up under independent evaluation — and nothing yet confirms they will — matters less than the cadence. The closed labs are now competing weekly with a model that anyone can download, and that is a race none of them signed up for.