# Precision & versioning
> What weights and numerical precision sit behind each Kurrens model id, and how that changes over time.
When you call a model on Kurrens, you should know exactly what you're running. We make three commitments.
## 1. The id is the Hugging Face repository
Model ids are Hugging Face ids — `deepseek-ai/DeepSeek-V4.1-Flash` is the model published at
[huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash). We serve the
developer's released weights with the developer's official chat template and tokenizer. We don't fine-tune, distill, or
merge them.
## 2. Precision is declared, never changed silently
Each model lists the numerical precision it runs at — for example `FP8`. The same value is published as `quantization`
in [kurrens.ai/models.json](/models.json) and shown on OpenRouter.
| Precision | What it means |
|---|---|
| `BF16` / `FP16` | The weights as the developer trained or released them. Highest fidelity, highest cost. |
| `FP8` | 8-bit floating point. Standard for serving large models; quality is very close to BF16 on most tasks. |
| `FP4` / `NVFP4` / `MXFP4` | 4-bit formats. Cheaper and faster; more quality loss on long reasoning and code. |
We never lower a model's precision behind an existing id. If we offer another precision, it is listed as a separate
model or [service tier](/docs/getting-started/pricing-and-billing).
## 3. New releases get a new id
When a developer publishes an updated model, it gets its own id — for example `deepseek-ai/DeepSeek-V4-Pro-0813` alongside
`deepseek-ai/DeepSeek-V4-Pro`. We don't swap the weights behind an id you already use. The older id keeps working until
it is retired under the [model lifecycle](/docs/models/lifecycle) policy, with at least 30 days' notice.
## Current catalog