Key Notes
- Alibaba released the open weights for Qwen3.8-Max, a 2.4-trillion-parameter Mixture-of-Experts model (~95B active per token, 1M-token context), alongside a much smaller Qwen3.8-27B, the first time Alibaba has open-sourced a flagship Max-class model rather than keeping it API-only. Hosted pricing is $2/M input and $6/M output, well below GPT-5.6 Sol and Fable 5.
- On Alibaba's own benchmarks it lands near the frontier, "second only to Fable 5," leading a few tests (PaperBench 93.0, IFBench 82.8) but trailing on harder ones (SWE-bench Pro 67.7 vs Fable 5's 80.0).
Alibaba has released the open weights for Qwen3.8-Max, a 2.4-trillion-parameter model it describes as one of the most capable available and “second only to Fable 5,” Anthropic’s top model. This follows the model’s July preview and marks the first time Alibaba has open-sourced a flagship Max-class system rather than keeping it behind an API, a notable shift from its practice with earlier Max models.
The release ships alongside a much smaller Qwen3.8-27B, giving developers a version that can run on modest hardware rather than requiring the enormous infrastructure the full model demands. Both use the Qwen3.8 architecture.
The full model is a sparse Mixture-of-Experts system with 2.4 trillion total parameters but only about 95 billion active per token, which keeps inference costs closer to a far smaller model while retaining the knowledge capacity of a large one. It offers a native context window that extends up to roughly one million tokens.
An important limitation applies to the open version: it is text-only and always reasons before answering. Users can tune the depth of that reasoning across low, medium and xhigh settings, but cannot fully disable it. Multimodal input, including image and video, and other advanced features remain available only through the paid Qwen3.8-Max API, a deliberate split between the free open weights and the commercial hosted product.
On price, the hosted API runs at about $2 per million input tokens and $6 per million output, well below GPT-5.6 Sol and Fable 5, continuing the pattern of Chinese labs undercutting US frontier pricing.
Near the Frontier, on Its Own Scorecard
Alibaba’s benchmark table places Qwen3.8-Max close to the top. It reports leading scores on some tests, including 93.0 on PaperBench and 82.8 on IFBench, and 86.6 on Terminal-Bench 2.1, ahead of Claude Opus 4.8 and Fable 5 but behind GPT-5.6 Sol.
On harder, more realistic coding benchmarks it trails the leaders, scoring 67.7 on SWE-bench Pro against Fable 5’s 80.0 and 73.5 on FrontierSWE against Fable 5’s 88.8. The generational jump over its own predecessor is large, with DeepSWE rising from 21.6 to 56.6.
Every one of these figures is self-reported, and no independent third-party benchmarks were available at release, a caveat that matters given Alibaba’s history of internal scores that outside testing did not fully confirm. Analysts also flagged two specific issues: the multimodal comparison table is measured against a weaker Qwen3.7-Plus rather than the stronger Max variant, flattering the improvement, and Alibaba’s own reinforcement-learning scaling curve peaks and then declines, hinting at diminishing returns.
The Open-Weight Squeeze Continues
The release is the latest move in a coordinated Chinese push to erode the advantage of closed US frontier models. It arrived within weeks of Moonshot’s Kimi K3, a 2.8-trillion-parameter open model that topped one frontend-coding leaderboard ahead of Fable 5, and DeepSeek’s V4 Pro, part of a run that has normalized the idea that a Chinese lab can sit within a few points of the US frontier while giving its weights away.
The strategic logic is consistent: rather than beat US models outright, undercut them on price and access enough that the discount becomes hard to refuse. Notably, all three models benchmark themselves specifically against Anthropic’s Fable 5, which has become the reference point defining the top of the market, even though its own access is tightly constrained by US export controls and usage limits.
For developers, the open weights are the real news, since they can run Qwen3.8 on their own servers, point existing OpenAI– and Anthropic-compatible coding tools at it without rebuilding, and avoid per-token cloud fees entirely. The deployable reality is nuanced: the hosted API is usable by any company today, but running the full 2.4-trillion-parameter weights requires serious infrastructure, which is why the smaller 27B checkpoint may matter more in practice. Whether Qwen3.8 genuinely narrows the gap with the US frontier, rather than on Alibaba’s own scorecard, awaits the independent testing the open weights now make possible.
Disclaimer: AIstify is an independent media brand owned and operated by NuvexMedia LLC, publishing news, research, and insights on artificial intelligence, emerging technologies, automation, and related industries. NuvexMedia LLC invests in and collaborates with companies across the AI, technology, software, and digital innovation sectors. These relationships do not influence AIstify’s editorial coverage, and the publication maintains full editorial independence to provide accurate, timely, and objective information. © 2026 NuvexMedia LLC. All rights reserved. This content is for informational purposes only and should not be considered legal, tax, investment, financial, or other professional advice.
