China AI Hub
Guides / Open-Weight vs API: Trade-offs for Chinese Models
Open-Weight vs API: Trade-offs for Chinese Models
Updated: 2026-09-22
The Chinese model ecosystem offers a genuine choice: some flagships are API-only, others ship open weights. This guide lays out the trade-offs using database fields (last verified 2026-09-22).
The two paths
- API: you pay per token, the provider runs everything. Qwen3.8-Max ($2.00/$6.00) and Doubao Seed 2.1 Pro ($6.00/$30.00) are API-only in our database.
- Open weight: you download the weights and serve them yourself. DeepSeek-V4-Pro (MIT), V4.1-Flash (MIT), GLM-5.3 (Apache-2.0), Kimi K3 (custom), MiniMax M3 (community) are all open weight.
When the API wins
- You need capability immediately and lack serving infrastructure.
- Your volume is low or bursty; fixed infrastructure costs would dominate.
- You want the vendor’s latest updates without re-deploying.
- You cannot satisfy the license obligations of the open-weight option.
At the database’s listed prices, the budget tier is aggressive: $0.15/1M input with $0.50-$0.60 output (GLM-5.3-Flash, DeepSeek-V4.1-Flash). Self-hosting rarely beats that below substantial, steady volume.
When open weights win
- Data residency or regulatory constraints require on-premises inference.
- Steady, predictable volume makes GPU costs amortize below API prices.
- You need to fine-tune or modify the model (licenses permitting).
- You need deterministic, offline capability.
Self-hosting economics: a 4-bit quantized 30B-class open model runs on a single high-memory GPU; frontier-scale open models need clusters. Quantization is the standard route — and its quality is your responsibility to evaluate.
Licenses are the fine print
“Open weight” is not one legal category in this ecosystem:
- MIT (DeepSeek V3.2/V4-Pro/V4.1-Flash): least restrictive.
- Apache-2.0 (GLM-5.x): permissive with patent grant.
- Kimi K3 License: MIT-style, but Model-as-a-Service operators above $20M revenue must sign a separate agreement; >100M MAU products must display the name.
- MiniMax Community License: non-commercial free; commercial use needs attribution plus authorization above revenue thresholds.
- Qwen3.8-2.4T-A95B: custom MIT-style with scale-triggered obligations.
A license that is permissive at your scale may not be at your customer’s scale. Review the exact text before shipping.
The hybrid reality
Most serious deployments run both: open-weight models for sensitive or steady workloads, APIs for spikes and frontier capability. The database fields to compare are open_weight, license, self_hosting, api_available and the pricing tables — all of which you can filter on the models page.
Bottom line
Choose the API when you need speed and simplicity; choose open weights when you need control, residency or fine-tuning — and read the license before you build.