Qwen 3.8-Max: Alibaba's most powerful model is now free to download

The Max tier was the part of Qwen nobody could download. Alibaba's flagship models stayed behind APIs while the smaller releases went open weight. Qwen 3.8-Max ends that. The release announcement calls it the most capable model in the Qwen family to date and the first Qwen-Max-class release with open weights, published through Hugging Face and ModelScope, a shift we flagged ahead of the release. It scales to 2.4 trillion total parameters, activating 95 billion per token, on the architecture Qwen 3.5 introduced.

It already has a track record from the closed era. In a 24-hour stretch, Qwen 3.8-Max beat 458 human teams working alone, as covered in the 24-hour contest report. Then came a 16-day autonomous run that produced an agent framework and a second-place ranking in Vision Arena, documented in the 16-day run coverage. Those capabilities are about to be reproducible anywhere the model can be served.

Alibaba's closed Max tier just went open source

This is a strategic break, not a goodwill gesture. Alibaba has consolidated its AI work into a single unit whose stated mission is to "create, deliver, and apply tokens," backed by a $53 billion investment push, per our breakdown of that spending. The company ships a fleet of models rather than a single champion, a platform play we mapped in the earlier strategy analysis, and the deal here is clear: the weights are free, while Qwen Cloud and the surrounding serving tooling are the paid channel. That paid channel now bundles text, vision, speech, and image models under one subscription, per the Qwen Cloud plan coverage. One quirk matters for anyone planning to run the flagship: the open 2.4T variant, Qwen3.8-2.4T-A95B, is text-only and requires thinking mode for every interaction. Multimodal inputs are unsupported and thinking cannot be disabled, even though the Qwen 3.5 foundation beneath it was trained as a unified vision-language model.

Where Qwen 3.8-Max beats GPT-5.6 Sol and where it loses

On the announcement's numbers, Qwen 3.8-Max clears its predecessor on nearly every row. Against closed rivals the wins are real but uneven. It takes top marks on PaperBench (93.0, ahead of GPT-5.6 Sol's 90.5), instruction following (IFBench 82.8 versus 72.7), and computer use (OSWorld-Verified 86.1), and it edges GPT-5.6 Sol on SWE-bench Pro, 67.7 to 64.6. Elsewhere the lead flips. Fable 5 stays well ahead on that same benchmark at 80.0, GPT-5.6 Sol tops Terminal Bench 2.1 at 88.8 against 86.6, and on GPQA Diamond the best number is GPT-5.6 Sol's 94.1.

Scores as published by the Qwen team. Per the announcement's footnotes, Fable 5 results may involve fallbacks.

Read the rows and a pattern appears: Qwen leads on instruction following, paper reproduction, and computer use, while repo-scale software engineering and hard reasoning still favor the closed models. The SWE-bench Pro baseline is also the team's own corrected version, with problematic tasks fixed and every competitor re-evaluated on the refined benchmark. That detail matters when the gap to GPT-5.6 Sol is only 3.1 points.

A 2.4T model that activates 95B parameters per token

The A95B suffix is the number that matters for anyone hosting it. The mixture-of-experts design activates 95 billion of the 2.4 trillion parameters per token, and that active count is what serving cost tracks. It is also what makes a model this size practical outside a lab; most parameters stay dormant. The Qwen 3.5 lineage contributes Gated Delta Networks alongside sparse MoE, which the team credits for high-throughput inference with minimal latency overhead. Context length reaches 1 million tokens, with recommended output budgets of 262,144 reasoning tokens and 131,072 final tokens. The same release includes a 27-billion-parameter sibling, Qwen3.8-27B, about 90 times smaller.

The efficiency pressure lands on the serving engines. The team points to SGLang, vLLM, or TokenSpeed for production workloads, the same territory Aleph Alpha's recent work targets: its megakernel library reported up to 200% inference speed gains for FP8 MoE models running in vLLM and SGLang. With weights this large now open, inference efficiency, not parameter count, decides who can actually use the model.

The harness caveat behind the agent gains

The agentic jumps are the headline: DeepSWE 1.1 climbs from 21.6 to 56.6 against Qwen 3.7-Max, OSWorld-Verified from 73.3 to 86.1, and the Agents' Last Exam pass rate from 11.8 to 27.0. The footnotes deserve as much attention as the scores. Qwen runs in many rows use the Claude Code harness, competitors get their official harnesses (GPT-5.6 Sol with Codex, Claude models with Terminus 2 scores from Artificial Analysis), and several benchmarks are in-house: QwenSWEBench, QwenQoderBench, CoWorkBench, and others. For SkillsBench, the team states plainly that all results are from its own testing. The Fable 5 numbers "may involve fallbacks."

That disclosure is unusually candid, but it is still vendor-run evaluation against closed rivals, on benchmarks the vendor partially controls. The gap between benchmark coding and production is not new: top models clear 96% on SWE-bench Verified yet barely reach 23% on private enterprise code, as our coverage of the Pro variants showed. Research we covered on harness evolution, tested on Terminal-Bench 2.1 with GPT-5.4 and Claude Opus 4.6, asks whether improvements found through harness iteration survive on held-out tasks. The same question hangs over Qwen's narrow SWE-bench Pro lead.

What developers can actually run with open weights

Two model IDs are published: Qwen/Qwen3.8-2.4T-A95B and Qwen3.8-27B, with downloads from Hugging Face or ModelScope. Users blocked from Hugging Face can switch via environment variables such as SGLANG_USE_MODELSCOPE or VLLM_USE_MODELSCOPE. The deployment layer has been catching up: Hugging Face reworked its Inference Endpoints around vLLM, SGLang, llama.cpp, and custom containers with autoscaling, which gives a 95B-active MoE a commodity hosting path without a bespoke stack.

The release extends Qwen's catalog strategy. Its Hugging Face presence spans seven capability domains, from language to speech, image, agent reasoning, and safety alignment, and the family keeps growing in quiet, off-mainstage releases. The team habitually ships training data, evaluation harnesses, and reward models alongside weights. Open-sourcing the flagship changes what that habit is worth: a model with top scores on PaperBench, IFBench, and OSWorld-Verified now comes with weights attached, something its closed rivals have not done.

The open question is whether the wins survive third-party verification. The footnotes show real candor about harnesses and in-house tests. But the rows where Qwen 3.8-Max beats GPT-5.6 Sol were measured by the team that trained the model, and the strongest prior independent check, Artificial Analysis on Qwen 3.7-Max, dates from the previous generation. Until independent runs land, the largest open-weight release Alibaba has ever shipped rests on numbers it largely recorded itself.