Two Chinese frontier models dropped within days of each other. Kimi K3 launched July 16 with 2.8 trillion parameters and open weights. Qwen 3.8 Max launched August 3 with 2.4 trillion parameters and multimodal capabilities. Both are competing for the title of best Chinese AI model.
The choice depends on what you need: open weights today, or multimodal capabilities.
The core difference
Qwen 3.8 Max is Alibaba’s flagship. 2.4T total params, 95B active (Sparse MoE), 1M context, multimodal (text + vision). It ranks #2 on Vision Arena and #4 on Frontend Code Arena. Open weights coming next week.
Kimi K3 is Moonshot AI’s flagship. 2.8T total params, ~200B active (estimated), 1M context, text-only. It scores 88.3% on Terminal-Bench (#2 overall). Open weights available now (1.56 TB on HuggingFace).
Specs comparison
| Spec | Qwen 3.8 Max | Kimi K3 |
|---|---|---|
| Total params | 2.4T | 2.8T |
| Active params | 95B | ~200B est. |
| Architecture | Sparse MoE + hybrid attention | MoE |
| Context window | 1M | 1M |
| Multimodal | Yes (text + vision) | No (text only) |
| Open weights | Next week | Available now |
| License | TBD | Open-weight |
| API price | Not published | $3/$15 |
Benchmarks
| Benchmark | Qwen 3.8 Max | Kimi K3 |
|---|---|---|
| Text Arena | #5 | not published |
| Vision Arena | #2 | N/A (text only) |
| Frontend Code Arena | #4 | not published |
| Terminal-Bench 2.1 | not published | 88.3% (#2) |
| Autonomous coding | 16 days | not published |
Direct comparison is limited because the models have been tested on different benchmarks. What we know:
- Qwen 3.8 Max leads on vision tasks (#2 Vision Arena) and frontend code (#4 Frontend Code Arena). It demonstrated 16-day autonomous coding.
- Kimi K3 leads on Terminal-Bench (88.3%, #2 overall). It is the largest open-weight model available today.
Open weights
Kimi K3: Available now. 1.56 TB on HuggingFace. You can self-host, fine-tune, and modify the model today.
Qwen 3.8 Max: Coming next week. When released, it will be the largest open-weight model. But you cannot self-host it yet.
If you need open weights today, Kimi K3 is the only option. If you can wait a week, Qwen 3.8 Max will be available.
Multimodal capabilities
Qwen 3.8 Max: Full multimodal support. Processes documents, video, and images natively. Vision Arena #2 ranking.
Kimi K3: Text only. No vision capabilities.
If you need multimodal processing (documents, images, video), Qwen 3.8 Max is the clear winner. If you only need text, both models are competitive.
Autonomous coding
Qwen 3.8 Max: 16-day autonomous coding demonstration. Built “oh-my-cli” from scratch over 16 days without human intervention.
Kimi K3: No comparable demonstration. Terminal-Bench score (88.3%) suggests strong coding capability, but no multi-day autonomous coding evidence.
For sustained autonomous coding, Qwen 3.8 Max has the stronger evidence.
When to use Qwen 3.8 Max
- You need multimodal capabilities (documents, images, video)
- You want 16-day autonomous coding capability
- You can wait a week for open weights
- You need Vision Arena-class performance
- You process visual content natively
When to use Kimi K3
- You need open weights today
- You want the largest active parameter count (~200B vs 95B)
- You need 88.3% Terminal-Bench performance
- You only process text (no vision needed)
- You prefer Moonshot AI’s ecosystem
My take
Both models are legitimate frontier contenders. The choice comes down to two factors: timing and modality.
If you need open weights today, Kimi K3 is your only option. The 1.56 TB weights are available on HuggingFace right now. Qwen 3.8 Max weights are coming next week, but “next week” in AI can mean anything.
If you need multimodal capabilities, Qwen 3.8 Max is your only option. Kimi K3 is text-only. The Vision Arena #2 ranking is impressive and matters for document processing, image analysis, and video understanding.
For pure text coding, both are strong. Kimi K3’s 88.3% Terminal-Bench is the highest among open models. Qwen 3.8 Max’s 16-day autonomous coding demonstration is the most compelling sustained agent evidence.
My recommendation: wait for the Qwen 3.8 Max open weights next week, then benchmark both on your specific tasks. The competition between these two models will drive improvements in both.
FAQ
Which is more powerful, Qwen 3.8 Max or Kimi K3?
They are competitive. Qwen 3.8 Max leads on vision (#2 Vision Arena) and autonomous coding (16 days). Kimi K3 leads on Terminal-Bench (88.3%, #2 overall). Direct comparison is limited by different benchmark suites.
Can I self-host Qwen 3.8 Max?
Not yet. Open weights are coming next week. When released, you will need multiple 80GB+ GPUs due to the 2.4T total parameter count. Kimi K3 weights are available now (1.56 TB).
Which is cheaper?
Kimi K3 at $3/$15 per 1M tokens. Qwen 3.8 Max pricing has not been published but is expected to be in the $3-$5/$10-$15 range.
Does Kimi K3 have vision capabilities?
No. Kimi K3 is text-only. If you need multimodal processing (documents, images, video), use Qwen 3.8 Max.
Which has better autonomous coding?
Qwen 3.8 Max has the stronger evidence: 16-day autonomous coding building a real project. Kimi K3 has no comparable demonstration, though its Terminal-Bench score (88.3%) suggests strong coding capability.
Should I wait for Qwen 3.8 Max open weights or use Kimi K3 now?
If you need open weights today, use Kimi K3. If you can wait a week, benchmark both when Qwen 3.8 Max weights drop. The competition will benefit you either way.