Kopro AI
Kopro AI
Our third generation, rebuilt on the Llama 3.2 architecture and tuned for natural conversation in English. Available today in Kopro Chat and the API.
Every previous Kopro model was built on Qwen. With 3.0 we moved the family to a new foundation: Meta's Llama 3.2, fine-tuned by our team for the way people actually use Kopro.
Kopro 3.0 is a 3.2-billion-parameter model, quantized to 4-bit (Q4_K_M) so the whole model fits in about 2 GB. It runs on our own CPU servers rather than on a third-party model provider.
We tuned it on a curated set of a thousand conversations covering everyday advice, science explanations, practical tech help, cooking, culture and creative ideation. The goal was not to win leaderboards but to make Kopro feel more natural to talk to.
Kopro 3.0 keeps the Llama 3.2 3B design: 28 transformer blocks, a 3,072-wide hidden state and grouped-query attention. Rotary position embeddings with a base of 500,000 let the architecture handle up to 128K tokens.
| Transformer blocks | 28 |
| Hidden size | 3,072 |
| Feed-forward size | 8,192 |
| Attention heads (Q / KV) | 24 / 8 |
| RoPE base | 500,000 |
| Vocabulary | 128,256 |
| Quantization | Q4_K_M · 2.0 GB |
We want to be upfront: Kopro 3.0 does not speak Russian well. Its Llama 3.2 base was trained mostly on English, and our fine-tuning did not change that. In Russian it makes grammar mistakes, loses meaning and sometimes switches to English mid-answer.
If you write in Russian, choose Kopro 2.0 in the model menu. It is built on Qwen 2.5, handles Russian far better and is about twice as fast on our servers.
We ran Kopro 3.0 and Kopro 2.0 through the same set of 14 evaluations shown on our home page: 11 auto-graded questions each, temperature 0, identical prompts.
On our CPU servers Kopro 3.0 writes more slowly than its siblings: it trades speed for a newer architecture. For quick back-and-forth, Kopro Flash is still the fastest choice.
Use the model ID kopro-3.0 with our OpenAI-compatible endpoint. Existing integrations only need to change one line.
from openai import OpenAI client = OpenAI(base_url="https://kopro.mom/v1", api_key="kp_live_...") reply = client.chat.completions.create( model="kopro-3.0", messages=[{"role": "user", "content": "Hi! Who are you?"}], ) print(reply.choices[0].message.content)
Open Kopro Chat and pick Kopro 3.0 from the model menu at the top.