Models
Each model is served exactly as listed: quantization, context and license are disclosed, not implied.
| Model | Parameters | Context served | Context native | Quant | License | Status |
|---|---|---|---|---|---|---|
Gemma 4 26B A4Bgoogle/gemma-4-26b-a4b-it | 25.2B total | — | 256K tokens | FP8 | Apache 2.0 | In testing |
Qwen3.8 27Bqwen/qwen3.8-27b | 27B | — | 256K tokens | FP8 | Apache 2.0 | In testing |
Both models support 256k tokens, but the context we can serve depends on how much GPU memory is left for the KV cache after the FP8 weights load. We list the measured figure, not the model card's.
About these models
Gemma 4 26B A4B
High-volume chat, extraction and classification where cost per request matters.
25.2B total; 3.8b active (mixture of experts).
Qwen3.8 27B
Coding and agent workflows with long prompts and optional reasoning.
27B; dense (all parameters active).
What FP8 means
FP8 stores each weight in 8 bits instead of 16. The model fits on one data-center GPU and runs faster, with a small effect on output quality. We state it because many providers do not, and we change this page before we change how a model is served.