Qwen 3.8 Max is here, and it’s a significant upgrade over Qwen 3.7 Max. The parameter count jumped from undisclosed (estimated 300B+) to 2.4 trillion. Multimodal support went from basic to Vision Arena #2. Autonomous coding went from hours to 16 days.
If you are using Qwen 3.7 Max, here is what changed and whether you should upgrade.
What changed
| Spec | Qwen 3.7 Max | Qwen 3.8 Max |
|---|---|---|
| Total params | ~300B est. | 2.4T |
| Active params | Not disclosed | 95B |
| Architecture | MoE | Sparse MoE + hybrid attention |
| Context window | 1M | 1M |
| Multimodal | Basic | Vision Arena #2 |
| Autonomous coding | Not demonstrated | 16 days |
| Open weights | No (API only) | Next week |
| API price | $2.50/$7.50 | Not published |
The biggest changes:
1. Parameter count: From ~300B estimated to 2.4T total (95B active). That is an 8x increase in total parameters. The active parameter count (95B) is also significantly larger.
2. Multimodal capabilities: Qwen 3.7 Max had basic multimodal support. Qwen 3.8 Max ranks #2 on Vision Arena. The vision capabilities are now a headline feature, not an afterthought.
3. Autonomous coding: Qwen 3.7 Max had no autonomous coding demonstration. Qwen 3.8 Max built “oh-my-cli” over 16 days without human intervention.
4. Architecture: Sparse MoE with hybrid attention is a new architecture. This enables the massive parameter count while keeping inference costs reasonable.
Benchmark comparison
| Benchmark | Qwen 3.7 Max | Qwen 3.8 Max |
|---|---|---|
| Text Arena | not published | #5 |
| Vision Arena | not published | #2 |
| Frontend Code Arena | not published | #4 |
| Terminal-Bench | not published | not published |
Qwen 3.7 Max did not publish Arena rankings. Qwen 3.8 Max’s rankings (#5 Text, #2 Vision, #4 Frontend Code) suggest significant improvement.
Pricing
Qwen 3.7 Max: $2.50/$7.50 per 1M tokens.
Qwen 3.8 Max: Not published. If priced similarly, it’s a free upgrade in capability. If priced higher (expected $3-$5/$10-$15), the cost increase may be worth it for the capability jump.
Should you upgrade?
Yes, if:
- You need multimodal capabilities (Vision Arena #2)
- You want autonomous coding (16 days)
- You process large codebases (1M context)
- You plan to self-host (open weights next week)
- The pricing is acceptable (TBD)
No, if:
- You are happy with Qwen 3.7 Max’s performance
- You need the lowest possible cost ($2.50/$7.50)
- You do not need multimodal capabilities
- Neither model’s open weights are available today (3.7 Max is API-only, 3.8 Max weights ship next week) and you need to self-host right now (consider Kimi K3 or DeepSeek V4 Pro instead)
Migration guide
Switching from Qwen 3.7 Max to Qwen 3.8 Max:
- Change model name from
qwen3.7-maxtoqwen3.8-max - Update SDK version (
pip install --upgrade dashscope) - Test your prompts (model behavior may differ)
- Monitor costs (pricing may differ)
- For multimodal: switch from
GenerationtoMultiModalConversationAPI
The API format is the same. No code changes needed beyond the model name.
What happens to Qwen 3.7 Max?
Alibaba has not announced deprecation. Qwen 3.7 Max will likely remain available for existing users. But new development will focus on 3.8 Max. If you are starting a new project, use 3.8 Max.
For existing projects, there is no urgency to upgrade. Qwen 3.7 Max still works fine. Upgrade when you need the new capabilities or when pricing makes sense.
Cost implications at scale
If you are running high volume through the API, the pricing gap matters more than the headline per-token numbers suggest. At Qwen 3.7 Max’s confirmed $2.50/$7.50, processing 10 million tokens (a moderate week of production traffic) costs roughly $25 input plus $75 output for a typical 1:1 in/out ratio, around $100 total. If Qwen 3.8 Max lands at the higher end of its expected $3-5/$10-15 range, the same workload could cost $130-200, a 30-100% increase. Budget for this uncertainty until Alibaba publishes official pricing, rather than assuming the upgrade is cost-neutral.
Should new projects start on 3.7 Max or wait for 3.8 Max?
If you’re starting a brand new project today rather than migrating an existing one, the calculus is slightly different than a straight upgrade decision. Qwen 3.7 Max is available now with confirmed pricing and a working integration path. Qwen 3.8 Max’s API is live but pricing is unconfirmed, and its open weights (the main reason to prefer it for a new self-hosted project) do not ship until next week.
For a project with a hard deadline this week, start on 3.7 Max and plan a migration once 3.8 Max’s pricing and stability are confirmed. For a project with more runway, it is reasonable to build against 3.8 Max from day one, especially if multimodal or long-horizon autonomous coding are core requirements rather than nice-to-haves.
The 8x parameter increase, Vision Arena #2 ranking, and 16-day autonomous coding are meaningful improvements. The question is whether you need them.
For most developers using Qwen 3.7 Max for text-based coding, the upgrade is optional. Qwen 3.7 Max at $2.50/$7.50 is still a good model. Qwen 3.8 Max adds capabilities you may not need.
For developers building multimodal applications or autonomous agents, the upgrade is essential. Vision Arena #2 and 16-day autonomous coding are capabilities Qwen 3.7 Max cannot match.
My recommendation: wait for pricing to be announced. If Qwen 3.8 Max is priced at $3/$10 or less, upgrade immediately. If priced at $5/$15, evaluate whether the new capabilities justify the 2x cost increase.
FAQ
Is Qwen 3.8 Max better than Qwen 3.7 Max?
Yes, significantly. 2.4T vs ~300B total params, Vision Arena #2, 16-day autonomous coding. The capability jump is substantial.
Will Qwen 3.7 Max be deprecated?
Not announced. Qwen 3.7 Max will likely remain available for existing users. But new development will focus on 3.8 Max.
How much more expensive is Qwen 3.8 Max?
Pricing not published. Qwen 3.7 Max costs $2.50/$7.50. Qwen 3.8 Max is expected at $3-$5/$10-$15. The cost increase depends on final pricing.
Can I use both models?
Yes. You can use Qwen 3.7 Max for simple tasks and Qwen 3.8 Max for complex tasks. This hybrid approach optimizes cost and capability.
Do I need to change my code?
Only the model name. Change qwen3.7-max to qwen3.8-max. The API format is the same.
When will Qwen 3.8 Max open weights be available?
Next week (announced August 3, 2026). When released, you can self-host on your own hardware.