⚡ Flash Models
Flash models trade context length for a much faster time-to-first-token. They are single-turn (every request starts from a clean KV state; a system prompt is pinned and reused), have a fixed 1k-token context, and do not support tool calling. They are built for high-throughput, short, independent requests: classification, captioning, extraction, routing, translation.
| Model tag | Task | Model card |
|---|---|---|
gemma4e-flash:e2b |
Any-to-Text (image, audio) | Gemma 4 E2B — Flash |
gemma4e-flash:e4b |
Any-to-Text (image, audio) | Gemma 4 E4B — Flash |
nemotron3.5-bolt-dflash-flash:e4b |
Any-to-Text (image, audio), speculative decoding | Nemotron Bolt E4B — DFlash Flash |
qwen3vl-flash:4b |
Image-Text-to-Text | Qwen3-VL-4B — Flash |
hy-mt2-flash:1.8b |
Text-to-Text (Translation) | Hy-MT2 |