Alibaba launched Qwen 3.8 Max — a 2.4T-parameter multimodal MoE model that just hit #4 on Frontend Code Arena with 1,668 points. At $2 per million tokens with open weights confirmed, it's the cheapest frontier-class coding model available.
Vamsi Tallapudi
Manager, Architect Technology at Cognizant
Alibaba just shipped Qwen 3.8 Max — a 2.4-trillion-parameter multimodal model that immediately landed at #4 on the Frontend Code Arena with 1,668 points. It costs $2 per million input tokens, the open weights are confirmed, and it's only one point behind Claude Opus 5 High. That's a lot of model for not a lot of money.
Qwen 3.8 Max was unveiled at the World AI Conference in Shanghai on July 19, 2026. It's Alibaba's biggest model yet — and their first multimodal model above one trillion parameters.
Here's what shipped:
The preview is live now on Alibaba Cloud, Qoder, and QoderWork. Alibaba also teased a smaller Qwen 3.8-27B variant for local deployment.
And then it hit the leaderboards.
Arena.ai confirmed the numbers. Here's the top of the Frontend Code Arena right now:
| Rank | Model | Score | Price (Input/Output) |
|---|---|---|---|
| #1 | Claude Opus 5 Max | 1,705 | $5/$25 per M tokens |
| #2 | Kimi K3 Max | 1,676 | $3/$12 per M tokens |
| #3 | Claude Opus 5 High | 1,669 | $5/$25 per M tokens |
| #4 | Qwen 3.8 Max | 1,668 | $2/$6 per M tokens |
Just 37 points separate first from fourth. And Qwen 3.8 Max is doing it at less than half the cost of either Claude model.
It's not just frontend either. Across the board:
That Vision result matters. Most coding models treat images as an afterthought. Qwen 3.8 processes them natively, which means you can feed it a design mockup and get working frontend code back — not just text-to-code, but image-to-code.
A lot, actually. The jump from 3.7 to 3.8 isn't incremental:
Alibaba claims it's "second only to Fable 5" overall. That's their internal evaluation, not independently verified — but the Code Arena and Vision Arena results give the claim some teeth.
This is what everyone's waiting for. Alibaba confirmed Qwen 3.8 Max will ship with open weights, which would make it the first Max-tier Qwen model you can actually download and run yourself.
But right now, it's a promise, not a download link. No date, no license, no HuggingFace repo. The model only runs through Alibaba's hosted APIs.
For context, Kimi K3 also promised open weights and delivered on July 27 — so there's recent precedent for Chinese labs following through. But there's also precedent for delays.
Depends on what you're optimizing for.
If you care about cost: This is the cheapest frontier-class coding model available. A typical session generating 100K output tokens costs about $0.60 on Qwen 3.8 Max, versus $2.50 on Claude Opus 5 and $6+ on GPT-5.6 Sol. For teams running hundreds of sessions monthly, that adds up fast.
If you care about quality: One point separating #3 and #4 is noise, not a real gap. On consumer product tasks, Qwen 3.8 actually beat both Claude Opus 5 variants.
If you care about ecosystem: This is where it gets tricky. Claude and GPT have mature IDE integrations, established enterprise agreements, and deep tooling. Qwen's API works fine, but the developer ecosystem around it is thinner. If you're already running Cursor or Claude Code, switching has friction.
If you need multimodal: Qwen 3.8 has a real edge here. Feeding a Figma screenshot and getting working React code back is something most competitors can't match at this price point.
The top of the Frontend Code Arena is now split across three countries — US (Anthropic), China (Moonshot, Alibaba) — with a 37-point spread across four models. The pricing gap between #1 and #4 is wider than the quality gap. A year ago, this board was US labs only. That's over.
New AI tools, automation workflows, and course drops — straight to your inbox. Join 2,400+ builders.
Alibaba launched Qwen 3.8 Max — a 2.4T-parameter multimodal MoE model that just hit #4 on Frontend Code Arena with 1,668 points. At $2 per million tokens with open weights confirmed, it's the cheapest frontier-class coding model available.
Elon Musk confirmed Grok 4.6 targets August 7 with 1.5T parameters and improved SFT/RL, followed weeks later by the 2.1T Grok 4.7. Here's the full xAI roadmap and why the split matters.
A leaked checkpoint reveals Cursor is testing Composer 3 internally under the codename Vega, with 4 reasoning tiers, 6 model variants, and performance reportedly between GPT-5.4 and GPT-5.5.
