Model benchmarks
Choose the AI that fits your slide
Decks are generated with our balanced default, but you can regenerate any single slide with a different model — right-click a slide in the rail, or use the Regenerate button under the preview. Here’s how each model performs on a standardized single-slide task. Measured on 2026-07-11.
Speed
Average response latency — lower is faster.
Throughput
Output tokens per second — higher is faster.
Reliability
Share of runs returning valid JSON.
| Model | Latency | Tokens/sec | JSON reliability | Context / Max out | Credit rate | Best for |
|---|---|---|---|---|---|---|
Llama 3.3 70BDefault Meta | 1.09s | 149 | 100% | 131k / 33k | 1.1× | Balanced default — reliable quality for any deck |
Llama 3.1 8B Instant Meta | 0.41s | 253 | 100% | 131k / 131k | 0.5× | Fastest & cheapest — quick drafts |
Llama 4 Scout 17B Meta | 0.64s | 220 | 100% | 131k / 8k | 1.2× | Fast with the most token headroom |
Qwen 3 32B Qwen | 1.16s | 361 | 100% | 131k / 41k | 2× | Strong reasoning, high throughput |
Qwen 3.6 27B Qwen | 1.87s | — | 60% | 131k / 33k | 2× | Multilingual prose (less reliable for strict JSON) |
GPT-OSS 20B OpenAI | 1.06s | 510 | 100% | 131k / 66k | 2.2× | Highest throughput — premium quality |
GPT-OSS 120B OpenAI | 1.16s | 291 | 100% | 131k / 66k | 2.5× | Largest open model — top quality |
Credit rate multiplies the token-based cost of a generation/regeneration. Regeneration is charged by tokens used × the model’s rate; a longer slide costs more than a short one. Numbers are averages and vary with load and prompt size.
New: regenerate any slide with a different model
- 1. Open your deck in the editor.
- 2. Right-click a slide in the left rail, or use the Regenerate slide button under the preview.
- 3. Pick a model — the slide’s text, tables, and charts are rewritten with fresh, real data.
- 4. Credits are charged by tokens used × the model’s rate shown above.