Which model should you plug into Cline, OpenCode, or Aider if you don't want to pay for tokens? We took every model that is currently free through an API in our catalog, kept those suited to coding, and ranked them by the independent Artificial Analysis Intelligence Index. Then we checked where each one is actually free and with what limits.
The ranking
| # | Model | AA Index | GPQA Diamond | Context | Free at |
|---|---|---|---|---|---|
| 1 | GLM 5.3 | 44.8 | 92% | 1M | NVIDIA Build, UnoRouter |
| 2 | Kimi K3 | 43.6 | 94% | 1M | NVIDIA Build, UnoRouter |
| 3 | GLM 5.3 Flash | 41.8 | 91% | 1M | NVIDIA Build, OrcaRouter, UnoRouter |
| 4 | Qwen 3.8 Flash Next | 39.8 | 92% | 1M | UnoRouter |
| 5 | DeepSeek V4.1 Flash | 39.5 | — | — | NVIDIA Build |
| 6 | DeepSeek V4 Pro | 36.0 | 93% | 1M | UnoRouter |
| 7 | DeepSeek V4 Flash | 34.3 | 91% | 1M | OrcaRouter, UnoRouter |
| 8 | Gemini 3.6 Flash | 34.0 | 93% | 1M | UnoRouter |
| 9 | Qwen3.8 27B | 33.7 | 91% | 262K | OpenRouter, UnoRouter |
| 10 | GLM 5.2 | 33.7 | 90% | 1M | OpenRouter, UnoRouter |
The Intelligence Index is a general score, not a pure coding benchmark, but it tracks agentic coding performance well because it includes coding and terminal tasks. Where Artificial Analysis publishes Terminal-Bench Hard for older versions, the ordering is similar: GLM 5.2 scores 50.8% and Kimi K2.6 43.9%.
What each model is good at
GLM 5.3 — the best free all-rounder
Z.ai's open-weight MoE model leads the free list. It handles multi-file edits and long agent runs well and keeps a 1M context on UnoRouter. On NVIDIA Build it is free for development use, which makes it the default choice for Cline or OpenCode.
Kimi K3 — strongest reasoning, image input
Moonshot's Kimi K3 scores the highest GPQA Diamond result among free models (94%) and accepts images, which helps when you debug from screenshots. Use it for planning and hard bugs, and a Flash model for routine edits.
GLM 5.3 Flash — fast and widely available
Almost as strong as the full model and available at three providers. If one route is rate-limited, you can switch to another with the same model ID family — ideal for long sessions.
Qwen 3.8 and DeepSeek V4 — solid alternatives
Qwen 3.8 Flash Next and the DeepSeek V4 line are strong at algorithmic code and refactoring. Qwen3.8 27B is small enough to also run locally in quantized form if you have a 24 GB GPU.
How to choose
- Agentic coding (Cline, OpenCode, Kilo Code): GLM 5.3 or Kimi K3 as the main model, GLM 5.3 Flash as fallback.
- Autocomplete: none of these — use a small fast model, ideally locally.
- Whole-repository questions: a 1M-context route: GLM 5.3 on UnoRouter or DeepSeek V4 Pro.
- Privacy-sensitive code: prefer local models; free routes may log prompts.
Setup instructions for seven tools are in How to use free LLMs in Cursor, Cline, OpenCode, and VS Code.
Methodology
We include text models that the catalog tags for coding (dedicated coder models plus strong general models known for code). Scores come from the public Artificial Analysis leaderboard and are matched automatically; where a model has several configurations, we show the best published one. We do not estimate missing scores. Availability is from our hourly snapshot of 15 providers.