Mellum2 Instruct

[!Note] Use this model when you want direct, low-latency answers without an explicit chain of thought — interactive chat, code assistance, tool use, and instruction following. If you need explicit reasoning before the answer (complex debugging, planning, multi-step agentic flows), use Thinking instead.

Mellum2 Instruct Highlights

Mellum2 Instruct is a post-trained 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 followed by reinforcement learning with verifiable rewards (RLVR) on math, executable coding, tool use, instruction following, reasoning, and knowledge tasks. Mellum2 Instruct answers directly, without an externalized chain of thought.

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 Instruct 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-Instruct --max-model-len 131072

# With tool calling
vllm serve JetBrains/Mellum2-12B-A2.5B-Instruct \
  --max-model-len 131072 \
  --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": "Write a Python function to reverse a string."},
]

chat_response = client.chat.completions.create(
    model="JetBrains/Mellum2-12B-A2.5B-Instruct",
    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 instruct (no-thinking) variants. All values are percentages; higher is better except HarmBench, where lower is better. All values self-reported by JetBrains.

Benchmark	Mellum2 Instruct SFT	Mellum2 Instruct	Qwen3.5 (4B)	Qwen3.5 (9B)	OLMo-3 (7B)	Ministral 3 (14B)	Seed-Coder (8B)
Coding
LiveCodeBench v6	30.9	37.2	51.0	63.7	28.2	42.4	28.1
EvalPlus	76.2	78.4	69.4	71.8	67.3	74.1	73.8
MultiPL-E	64.6	67.1	51.0	67.1	36.1	71.5	77.0
Tool Use
BFCL v4	31.8	44.2	52.0	60.6	19.8	38.8	—
BFCL v3	43.1	66.3	64.1	70.5	41.9	52.7	—
Math
AIME	29.9	41.7	38.3	58.3	40.0	33.3	0.0
GSM-Plus	73.0	80.5	85.2	87.9	85.8	86.6	50.4
Knowledge
MMLU-Redux	77.4	78.1	87.5	91.1	71.8	85.9	38.1
GPQA Diamond	38.9	40.9	76.8	79.8	40.9	58.6	20.2
Conversational
IFEval	69.3	75.8	82.1	83.9	83.2	67.3	56.2
JetBrains pairwise	66.7	68.1	60.6	77.8	44.4	72.4	43.0
MixEval	62.9	62.2	65.9	71.1	59.4	71.2	37.2
BS-Bench	24.0	18.0	56.9	61.0	22.0	9.0	5.0
Safety
HarmBench (↓)	8.4	23.1	20.3	20.9	14.7	56.5	40.0
XSTest	78.3	81.2	93.2	91.2	91.2	96.8	86.3

Notes:

EvalPlus is the mean of HumanEval+ and MBPP+.
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.

For more details, see the Mellum2 Technical Report.

License

Released under the Apache 2.0 license.