Alibaba's 2.4T Flagship Is Live. The Weights Walk Free Next Week.

TL;DR

Alibaba's Qwen team officially released Qwen3.8-Max today: a 2.4-trillion-parameter mixture-of-experts flagship with 95B active parameters, a 1M-token context window, image, text, and video input, and API access on QwenCloud at $2 per million input tokens and $6 per million output. The headline for the open-model crowd sits in the first paragraph of the announcement: this is "the first time we will open-source the weights of a Qwen-Max-class model," with the drop promised on Hugging Face and ModelScope next week. The benchmark table is unusually honest, showing losses to Claude Fable 5 on SWE-bench Pro and FrontierSWE next to wins on PaperBench and Terminal-Bench, and nearly all of it was measured inside Anthropic's own Claude Code harness.

Sixteen days from one tweet to 5,068 words

When Qwen 3.8 was announced on July 19, the entire release consisted of a single tweet: 2.4T parameters, a claim of being second only to Fable 5, and no benchmarks, no model card, no active-parameter count. Today's launch is the opposite: a 5,068-word post, a benchmark table with roughly 60 rows, public demo repositories, and pricing. Somebody found the numbers.

The disclosure that matters most is the sparsity. Pre-release coverage repeatedly flagged that Alibaba would not say how many parameters were active; the answer turns out to be 95B of 2.4T, roughly 4% of the model firing per token. Think of a consulting firm with 2.4 trillion specialists on the org chart but only 95 billion on call for any given question: you pay compute for who shows up, not the whole payroll. You do, however, still need a building big enough to house everyone, which is why the promised weights will matter mostly to clusters, not desktops.

What actually shipped today

• API on QwenCloud now: $2/$6 per million tokens, with cached input at $0.25. Endpoints in Beijing, Singapore, and Virginia speak both OpenAI-compatible and Anthropic-compatible protocols.
• 1M-token context (roughly 991.8K max input, 131K max output), with image, text, and video input.
• Adjustable reasoning: a reasoning_effort parameter with low, medium, and xhigh tiers. Note for your bill: xhigh is the default, and preserve_thinking is on by default too.
• Open weights next week, on Hugging Face and ModelScope, the first Qwen-Max-class model to get them. No license named yet.

The table Alibaba published anyway

Vendor benchmark tables usually only contain wins. This one does not, and that is the most interesting thing about it. On SWE-bench Pro, Alibaba's own numbers put Qwen3.8-Max a full 12 points behind Claude Fable 5, and on FrontierSWE the gap is 15:

Then there are the wins. On PaperBench, OpenAI's own benchmark for reproducing AI research papers, Qwen3.8-Max posts the top score, above GPT-5.6 Sol and both Claudes:

On Terminal-Bench 2.1 it reports 86.6, above both Opus 4.8 and Fable 5 (84.6 each, per Artificial Analysis) but under GPT-5.6 Sol's 88.8. The pattern across the table: second tier on pure software engineering, first tier on research reproduction and instruction following (82.8 on IFBench, ten points clear of GPT-5.6 Sol), and genuinely dominant on the multimodal-agent sheet, where it takes OSWorld-Verified at 86.1 and posts an 87.0 on the Dense200 counting benchmark against Fable 5's 31.1. If your agent needs eyes, this table says Qwen is now the model to beat.

Benchmarked in Anthropic's CLI

Read the footnotes and a pattern emerges: SWE-bench Pro, DeepSWE, NL2Repo, FrontierSWE, WideSearch, and both in-house coding suites were all evaluated with Claude Code as the harness. One footnote states it outright: "Qwen3.8-Max performs best on Claude Code." Alibaba also says it RL-trained the model across five harnesses (QwenWork, Claude Code, Codex, OpenClaw, Hermes), and the launch post ships copy-paste instructions for pointing Claude Code at QwenCloud via an Anthropic-compatible endpoint:

export ANTHROPIC_MODEL="qwen3.8-max"
export ANTHROPIC_BASE_URL=https://dashscope-intl.aliyuncs.com/apps/anthropic
export ANTHROPIC_AUTH_TOKEN=your_api_key
claude

There is something quietly remarkable about a frontier lab publishing its flagship's report card as measured inside a competitor's product, then recommending that competitor's product as the way to use it. The agent harness, not the model API, is becoming the neutral ground where everyone meets.

The stunt reel

The announcement's centerpiece demos are long-horizon autonomy runs, and Alibaba published traces for the biggest one. In a roughly 16-day fully autonomous run, the model built and maintained a CLI project in public at qwen-code-dev-bot/oh-my-cli, accumulating 265 commits, 127 pull requests, and 151 issues through its own issue-dispatch state machine, with no human in the loop.

The chip-design demo is the most concrete. Given empty Verilog templates and a cocotb test bench, the model spent about 500 turns iterating on a GCD/RSA accelerator through an edit-simulate-synthesize loop with Yosys and OpenROAD, taking its first working design from 8,298 gates down to 678 while holding bit-exact correctness, then closing timing at 500 MHz in physical layout with an 81% smaller die:

The rest of the reel: entered live in a Tianchi multimodal-intent competition against 526 human teams, it climbed from 0.60 to 0.853 accuracy over 45 submissions in 24 hours, beating 458 of them. In a 365-day e-commerce simulation it turned ¥100,000 into ¥416,252, which Alibaba says beats second-place GLM 5.2 by 38%. These are vendor demos on vendor-picked tasks, but publishing the full oh-my-cli trace is more receipts than these launches usually come with.

The fine print

Every number above is Alibaba's own, hours old, with no independent replication yet. The footnotes deserve a read: the SWE-bench Pro runs used a "refined" version of the benchmark with problematic tasks corrected, several rows mix Alibaba's in-house measurements with other vendors' best published scores, and the Fable 5 results carry a note that they "may involve fallbacks." The open-weights commitment is real text in an official post, but so far it is a date-shaped promise from the company whose Qwen-Image-3.0 shipped with no weights at all, and next week's release has no named license. Moonshot's Kimi K3 already proved a multi-terabyte frontier drop can actually land; whether Alibaba matches that, and under what terms, is the thing to watch.

Key Takeaways

• Qwen3.8-Max is live on QwenCloud: 2.4T total parameters, 95B active, 1M context, image/text/video input, $2/$6 per million tokens, reasoning_effort tiers with xhigh as default.
• Alibaba says the weights land on Hugging Face and ModelScope next week, the first open-weight Qwen-Max-class model. License still unannounced.
• The vendor's own table is candid: it trails Claude Fable 5 by 12.3 points on SWE-bench Pro and 15.3 on FrontierSWE, while winning PaperBench (93.0), Terminal-Bench 2.1 vs both Claudes (86.6), IFBench (82.8), and most of the multimodal-agent sheet.
• Nearly all coding evals ran inside the Claude Code harness, the model was RL-trained across five harnesses including Claude Code and Codex, and the launch post documents an Anthropic-compatible API for exactly that setup.
• Flagship demos are long-horizon: a 16-day autonomous run with public traces (265 commits, 127 PRs), and a 500-turn chip-design loop that cut a verified design from 8,298 to 678 gates.
• All numbers are vendor-reported and hours old; wait for independent evals, and treat "next week" as a promise until the weights are on the Hub.

Sources: Qwen official announcement, QwenCloud model page, @Alibaba_Qwen launch tweet, oh-my-cli autonomous-run repository, FrontierSWE leaderboard, Artificial Analysis