Cline – which open weight models to use
Short version: use a cheap, fast model for most coding and save the expensive ones for hard problems. My picks come from benchmark roundups that mix independent results with vendor-reported numbers, so treat them as a starting point.
Day-to-day coding: DeepSeek V4.1 Flash. In one agentic coding benchmark it scored 80% overall at about $0.14 per long session, against 72% for Kimi K3 and K3.1 Pro. It was also the fastest of the group. That matters on a subscription, because cheaper models stretch your quota further. The subscription meters usage against reference token prices. Best open-weight model for coding: Kimi K3.1 Pro vs GLM 5.3 vs DeepSeek V4.1 Flash +2
Hard or long-running coding: GLM-5.3. Z.ai positions it for long-horizon agentic coding and complex project delivery. It scored 74% in the same benchmark, second to DeepSeek. It’s text-only, so use GLM-5.3-Flash if you need image input. you
Frontend and UI work: Kimi K3. It’s #1 on the Arena.ai Frontend Code Arena and accepts images. It costs much more per session in that benchmark (about $2.46 against $0.14), so it burns quota faster. morphllm
Strategy and planning: I couldn’t find solid benchmarks aimed at planning rather than coding. My guess is you want a high-reasoning model for it:
- GLM-5.3: reasons on every request.
- Kimi K3: also reasons on every request.
- Qwen3.8 Max: a good general-purpose option, described as strong on multilingual work, agents and multimodal input. geotoolbox
Try the same planning prompt on all three and see which you prefer.
Probably skip: DeepSeek V4 Pro, Qwen3.7 Plus/Max, MiMo-V2.5 and MiniMax-M3 look like the previous generation. Muse Spark 1.3 Contributor I couldn’t find enough on to judge.
Many of these scores are self-reported by the labs, and short-task results don’t always hold up over long agent sessions. A practical approach is to run DeepSeek V4.1 Flash by default, switch to GLM-5.3 or Kimi K3 when it gets stuck, and see how fast your quota drains.