Free plan: free open models only
Free uses OpenRouter models that cost nothing to call. If one is busy or rate limited, Keplar tries another free model; it never falls back to a paid one. Free answers can be slower when free models are busy, and the hosts of free models may log prompts, so do not paste anything you would not want logged. Free allows up to 50 questions a day (fair use).
| Model | Provider | Used as | Weights | Context (tokens) | Reads images |
|---|---|---|---|---|---|
| Nemotron 3 Super | NVIDIA | Panel | Open weights | 262k | No |
| Gemma 4 31B | Panel | Open weights | 262k | Yes | |
| Qwen3.8 27B | Alibaba Qwen | Panel | Open weights | 262k | Yes |
| Dots3-Note Preview | Dots Studio | Panel | Open weights | 512k | Yes |
| Gemma 4 26B A4B | Verifier | Open weights | 262k | Yes | |
| Nemotron 3.5 Lightning | NVIDIA | Fast or fallback | Open weights | 1M | No |
| Nemotron 3 Ultra | NVIDIA | Synthesizer (judge) | Open weights | 1M | No |
| Nemotron 3 Nano Omni | NVIDIA | Fast or fallback | Open weights | 256k | Yes |
Paid plans: premium models added
Lite, Plus and Pro add the roster below. Plans differ in credits and limits, not in which question gets which model; see pricing for the current plans and what each includes.
| Model | Provider | Used as | Weights | Context (tokens) | Reads images |
|---|---|---|---|---|---|
| Gemini 3.1 Flash-Lite | Panel | Hosted only | 1.05M | Yes | |
| DeepSeek V4 Flash | DeepSeek | Panel | Open weights | 1.05M | No |
| Qwen3.8 Flash | Alibaba Qwen | Panel | Hosted only | 1M | Yes |
| Mistral Small (2603) | Mistral | Panel | Hosted only | 262k | Yes |
| Gemini 3.8 Flash | Synthesizer (judge) | Hosted only | 1.05M | Yes | |
| GPT-6 Luna | OpenAI | Panel | Hosted only | 1.05M | Yes |
| GPT-5 nano | OpenAI | Fast or fallback | Hosted only | 400k | Yes |
| Nemotron 3 Super 120B | NVIDIA | Panel | Open weights | 262k | No |
| GLM-5.3 Flash | Zhipu / Z.ai | Verifier | Open weights | 1.31M | Yes |
| Gemma 4 31B | Panel | Open weights | 262k | Yes | |
| gpt-oss-20b | OpenAI (open weights) | Fast or fallback | Open weights | 131k | No |
How Keplar picks models for a question
- A simple question goes to one model. A harder one goes to a panel of models from different families, plus a verifier.
- The router weighs quality, reliability, speed and cost, and tries to include models from different providers so the panel does not share one set of habits.
- Questions with images or video only go to models that can read them.
- A model that times out or refuses is not counted as disagreeing; the answer shows how many responded.
How the comparison and synthesis work is on How Keplar works.
What this page does not tell you
- It does not rank the models or claim one is better. Keplar has not published its own benchmark of them.
- It lists the roster, not what each question used. Open Models consulted under any answer to see that.
- Premium providers such as Anthropic's Claude models are not listed because they are not in the current live roster.