Models for fine-tuning in Nebius Token Factory

Nebius Token Factory supports fine-tuning on multiple open-weight model families.
This page lists:

• Which base models you can fine-tune
• Which context lengths they support
• Which fine-tuning types are available (LoRA vs full fine-tuning)

---

​

Model List

For each models listed below, Nebius Token Factory supports the following

context_length: 8192, 16384, 32768, 65536, 131072 Unless you override it via the context_length hyperparameter, the default context length for fine-tuning is 8192 tokens. Check hyperparameter section for model details regarding context_length

---

​

DeepSeek

Nebius Token Factory integrates DeepSeek V3 and V4 models for high-capacity reasoning workloads.
DeepSeek and its variants are released under the MIT License (see model cards for details).

---

​

Google

Nebius Token Factory supports full fine-tuning for select Google Gemma 4 instruction-tuned models. These models are available under the Apache 2.0 License. Review the corresponding model card for full license terms and usage details.

---

​

OpenAI / Unsloth GPT-OSS

These models are OpenAI GPT-OSS weights (bf16) packaged by Unsloth.
They are Apache 2.0–licensed and suitable for both research and commercial use (subject to the license). To convert the weights into MXFP4 please follow instructions here.

---

​

Qwen

Nebius Token Factory supports dense, moe and coder variants across Qwen3 and Qwen2.5 families.
All Qwen models below use the Apache 2.0 license (see each model card for details).

​

Qwen3 MoE

​

Qwen3 coder

​

Qwen3 dense + base

​

Qwen3.5

​

Qwen3.6

​

Qwen2.5 dense + coder

---

​

Meta (Llama 3.1 / 3.2 / 3.3)

Nebius Token Factory and the Meta models hosted in the service are built on the Llama 3.1, Llama 3.2, and Llama 3.3 families. For acceptable use, see Meta’s policies:

• Llama 3.1 Acceptable Use Policy
• Llama 3.2 Acceptable Use Policy
• Llama 3.3 Acceptable Use Policy

Deployment options currently only include via Dedicated endpoints.