Image: EXAI Global / YouTube Alibaba's Qwen3.8-Max Is Out of Preview: 2.4T Parameters, Competitive Coding Benchmarks, Open Weights Coming
Alibaba's biggest model yet went from preview to general availability on August 3. Qwen3.8-Max is a 2.4-trillion-parameter mixture-of-experts model with a 1M-token context window and strong agent benchmarks — though independent evaluations haven't caught up yet.
Alibaba’s Qwen3.8-Max moved from preview to general availability on August 3, 2026. The model is a 2.4-trillion-parameter mixture-of-experts architecture — the largest the Qwen team has shipped — with 95 billion active parameters per request, a 1M-token context window (991K input, 131K output), and native multimodal support for text, images, and video.
It’s available now via OpenAI-compatible and DashScope APIs. Open weights for Qwen3.8-Max and a smaller 27B checkpoint are promised for the week of August 10.
What the benchmarks say
Alibaba’s own benchmark table shows a model that’s competitive with frontier coding tools on agent-style tasks, but still trails the top tier on pure coding benchmarks:
- Terminal-Bench 2.1: 86.6 — ahead of Claude Opus 4.8 at 84.6, behind GPT-5.6 Sol at 88.8
- SWE-bench Pro: 67.7 — behind Claude Fable 5’s 80.0
- FrontierSWE: 73.5 — behind Fable 5’s 88.8
- GPQA Diamond (reasoning): 92.6
- OSWorld-Verified: 86.1
The pattern: the model is strong on agentic and terminal tasks, and particularly competitive on multimodal benchmarks. On pure coding tasks it doesn’t match the current leaders.
One important caveat — as of August 3, no independent evals exist. Artificial Analysis and community leaderboards hadn’t published numbers at launch. Alibaba’s table compares Qwen3.8-Max against other models, but those comparisons should be treated as self-reported until third parties run the same tests.
Pricing
Hosted API pricing is $2.00 input / $6.00 output per million tokens, with cached input at $0.25 per million tokens. The 8x cache discount is significant for agentic workflows with long shared contexts.
Context: how this fits
Qwen3.8-Max is Alibaba’s response to a stretch of China-based model releases — Moonshot’s Kimi K3 (2.8T open weights, now in GitHub Copilot), DeepSeek’s Flash model, and others. The competitive framing is mostly about reasoning benchmarks, where Alibaba claims parity with Anthropic. Coding-specific results are closer to the second tier.
The open-weight release scheduled for next week matters more for the developer community than the API launch. Open weights mean the model can be self-hosted, fine-tuned, and benchmarked independently — which is when a clearer picture of where it actually stands on coding tasks will emerge.
Sources: MarkTechPost · Bloomberg · Wan27 benchmarks