[!Note] Use this model when you want explicit chain-of-thought before the final answer — complex debugging, multi-step planning, agentic workflows, and math- or reasoning-heavy tasks. For direct, low-latency answers without reasoning traces, use Instruct instead.

Mellum2 Thinking Highlights

Mellum 2 Thinking is a post-trained reasoning-augmented assistant model trained by JetBrains.

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.

It is produced from Mellum2-12B-A2.5B-Base by supervised fine-tuning (loss computed only on the final assistant turn) followed by reinforcement learning with verifiable rewards (RLVR) on a harder data mix that includes a long-form math subset. The model emits its reasoning inside <think>...</think> blocks before the final answer.

Mellum2 Model Family

This repository contains one checkpoint from the Mellum 2 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 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
License: Apache 2.0

Serving with vLLM

# Without tool calling
vllm serve JetBrains/Mellum2-12B-A2.5B-Thinking \
  --max-model-len 131072 \
  --reasoning-parser qwen3

# With tool calling
vllm serve JetBrains/Mellum2-12B-A2.5B-Thinking \
  --max-model-len 131072 \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice \
  --tool-call-parser hermes

Quickstart

Text-Only Input

python

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",
    messages=messages,
    max_tokens=81920,
    temperature=0.6,
    top_p=0.95,
    extra_body={
        "top_k": 20,
    },
)
print("Chat response:", chat_response)

Evaluation

Post-training evaluation for the thinking/reasoning variants. All values are percentages; higher is better except HarmBench, where lower is better. All values self-reported by JetBrains.

Benchmark	Mellum2 Thinking SFT	Mellum2 Thinking	Qwen3.5 (4B)	Qwen3.5 (9B)	OLMo-3 (7B)	Ministral 3 (14B)
Coding
LiveCodeBench v6	75.1	69.9	59.4	68.3	59.8	42.7
Tool Use
BFCL v4	38.8	45.6	42.9	42.7	—	35.9
BFCL v3	60.5	69.4	73.9	68.5	—	52.2
Math
AIME	20.0	58.4	68.3	73.4	61.7	38.3
GSM-Plus	62.6	87.0	89.3	90.7	88.1	86.5
Knowledge
MMLU-Redux	84.8	86.2	88.3	91.7	71.3	84.4
GPQA Diamond	39.9	57.6	76.8	81.3	29.3	46.0
Conversational
IFEval	69.1	76.5	87.1	89.8	84.7	59.7
JetBrains pairwise	64.4	69.5	40.5	56.7	32.2	63.8
MixEval	63.4	66.9	71.9	76.0	67.0	70.8
BS-Bench	14.0	15.0	63.0	70.0	23.0	9.0
Safety
HarmBench (↓)	12.2	20.6	15.9	6.6	48.7	70.0
XSTest	90.8	89.6	96.8	97.6	93.2	96.8

Notes:

AIME is the mean of AIME 2025 and AIME 2026 (30 questions each).
BFCL v4 is the macro-average of five subtasks: v1, v2, v3, web search, memory.
JetBrains pairwise is win rate against Qwen2.5-7B-Instruct on an internal benchmark.
— indicates the model lacks native tool calling (OLMo-3-7B-Thinking).

For more details, see the Mellum2 Technical Report.

License

Released under the Apache 2.0 license.