How to choose a Qianfan model for your workload
2026-11-13
How to choose a Qianfan model for your workload
The Qianfan platform currently serves 17 models on the international edition
(snapshot: 2026-11-12). Reading the list without a frame wastes hours; here is
the frame we use internally.
First, separate the axes
- Capability tier: the "flash / light" models are latency and cost plays;
the "pro" models are reasoning plays. A task that fails on a flash model will
not be fixed by a bigger context window.
- Context window: 131k vs 202k vs 1M tokens. Most "long context" work is
actually "long context + retrieval": chunk it first, and you rarely need the
1M window at all.
- Modality: text-only models plus a separate image/video path. If your task
produces pixels or frames, you are shopping in a different catalog entirely.
A routing rule that survives contact with reality
1. Start every workload on the cheapest flash model that plausibly works.
2. Keep one pro model in your back pocket for the tasks the flash model fails,
and measure the failure rate rather than assuming it is zero.
3. Pin model IDs in config, never in prompt text — model IDs are point-in-time
snapshots, and the console is the source of truth for what exists today.
The dated snapshot on the homepage lists the current
catalogue with published per-token prices. The gateway does not rewrite the
model field: whatever you send is what the upstream receives, which is the
whole point of a proxy you can audit.
The failure mode nobody budgets for
Switching models mid-project. If your prompt has vendor-specific formatting
baked in, your "model-agnostic" code is a lie. Keep the agent logic and the
model adapter in separate files from day one; the second model swap will take
minutes instead of a week.
For the structured contract behind this advice, see the
API reference, and for how framework choice interacts
with model choice, see picking an agent framework.