Mellum2 Thinking (SFT)
[!Note] Research artifact: the SFT-only intermediate checkpoint of the Thinking training pipeline. Use it to study or build on the reasoning SFT stage in isolation, run post-training experiments (preference tuning, RLVR, etc.), or compare against the final RL-tuned variant. For production-quality reasoning use Thinking instead.
Mellum2 Thinking (SFT) Highlights
Mellum2 Thinking (SFT) is the supervised-fine-tuning-only intermediate checkpoint of the Mellum2 Thinking training pipeline, released by JetBrains as a research artifact.
The model uses a Mixture-of-Experts architecture with 64 experts and activates 8 experts per token. It uses a combination of sliding-window and full attention layers, with a context length of 131,072 tokens.
This checkpoint was produced from Mellum2-12B-A2.5B-Base with three epochs of SFT on a corpus of reasoning-trace-bearing data. The model emits its reasoning inside <think>...</think> blocks. It is the starting point of the RLVR stage that produces the final Mellum2-12B-A2.5B-Thinking.
Mellum2 Model Family
This repository contains one checkpoint from the Mellum2 family.
| Checkpoint | Description |
|---|---|
| Base Pretrain | Base checkpoint before long-context extension |
| Base | Final base model |
| Instruct SFT | Supervised instruction-tuned checkpoint |
| Thinking SFT | Supervised thinking checkpoint |
| Instruct | RL-tuned instruction model |
| Thinking | RL-tuned thinking model |
Model Overview
Mellum2 Thinking (SFT) has the following features:
- Number of Layers: 28
- Hidden Size: 2304
- Intermediate Size: 7168
- MoE Intermediate Size: 896
- Number of Experts: 64
- Number of Activated Experts: 8
- Number of Attention Heads (GQA): 32 for Q and 4 for KV
- Context Length: 131,072
- Sliding Window: 1,024
- Vocabulary Size: 98,304
- Precision: bfloat16
Serving with vLLM
# Without tool calling
vllm serve JetBrains/Mellum2-12B-A2.5B-Thinking-SFT \
--max-model-len 131072 \
--reasoning-parser qwen3
# With tool calling
vllm serve JetBrains/Mellum2-12B-A2.5B-Thinking-SFT \
--max-model-len 131072 \
--reasoning-parser qwen3 \
--enable-auto-tool-choice \
--tool-call-parser hermes
Quickstart
Text-Only Input
from openai import OpenAI
# Configured by environment variables
client = OpenAI()
messages = [
{"role": "user", "content": "Is 1024 a power of 2? Explain your reasoning."},
]
chat_response = client.chat.completions.create(
model="JetBrains/Mellum2-12B-A2.5B-Thinking-SFT",
messages=messages,
max_tokens=81920,
temperature=0.6,
top_p=0.95,
extra_body={
"top_k": 20,
},
)
print("Chat response:", chat_response)
Evaluation
Evaluation results are available in the model card. All values are self-reported by JetBrains.
For more details, see the Mellum2 Technical Report.
License
Released under the Apache 2.0 license.