Three Cheap, Capable Open-Weight Models
DeepSeek, Kimi (Moonshot), and MiniMax are the three open-weight Chinese LLMs most Western developers evaluate for cost-efficient production. They overlap a lot on price, but differ on the things that actually decide a build: context, multimodality, and what each was trained to do. All three are on LinkModel under one key.
The Specs
| DeepSeek V4 Flash | Kimi K2.6 | MiniMax M2.7 | |
|---|---|---|---|
| Input / Output (1M) | $0.14 / $0.28 | $0.95 / $4.00 | $0.30 / $1.20 |
| Cached input | ~$0.0028 (98%) | $0.16 (~6x) | — |
| Context | 1M | 256K | 205K |
| Multimodal | Text | Text + image | Text |
| License | MIT | Modified-MIT | Modified-MIT |
| Trained for | Efficiency, cache-heavy | Multimodal agent swarms | Long-horizon agents |
Where Each Wins
- DeepSeek V4 Flash — the cost king. Cheapest by far, 1M context, and a 98% cache discount that makes cache-heavy workloads (RAG, chatbots, agents with stable prefixes) nearly free on input. Default for high-volume, stable-context, output-heavy work. (Step up to V4 Pro for competition-grade coding.)
- Kimi K2.6 — the only one of the three that takes image input. Pick it for UI-from-screenshot, document-with-figures, and large multi-agent orchestration (up to 300 sub-agents). Beats GPT-5.4 on HLE-Full with tools.
- MiniMax M2.7 — the cheapest capable agent model, explicitly trained for multi-agent collaboration, live debugging, and multi-step delivery. Best when the model must hold a coherent plan across dozens of tool calls over long runs.
Quick Decision
| Need | Pick |
|---|---|
| Cheapest / cache-heavy / high-volume | DeepSeek V4 Flash |
| Image input / UI generation | Kimi K2.6 |
| Long-horizon agents on a budget | MiniMax M2.7 |
| Competition-grade coding | DeepSeek V4 Pro |
Shared Caveats
- Data residency: all three official APIs process data in China (a GDPR consideration). Self-host or use a zero-retention gateway. LinkModel defaults to zero data retention.
- Licensing: DeepSeek is fully MIT; Kimi and MiniMax are Modified-MIT (commercial self-hosting needs authorization).
- Thinking modes bill at the same rate but burn more tokens — off by default.
The Best Answer Is "Route"
These aren't mutually exclusive. A common stack: DeepSeek V4 Flash as the cheap default, MiniMax M2.7 for long agent plans, Kimi K2.6 when a task has images. Because all three are OpenAI-compatible and share one request shape on LinkModel, routing is a config change. More in open-source LLM API and best coding LLM API.
Start free with a $1 credit and test all three on your workload.