Description
Mellum2.1 Thinking is an advanced large language model developed by JetBrains, hosted on Hugging Face. It is designed to handle complex agentic tasks such as working in repositories, running commands, and calling tools, as well as tackling hard non-agentic problems in coding, math, and reasoning. The model is a 12 billion parameter mixture-of-experts model with 2.5 billion active parameters, featuring 28 layers, a hidden size of 2304, and a vocabulary size of 98,304.
Read More
The architecture of Mellum2.1 remains unchanged from its predecessor, Mellum2, but significant improvements have been made in post-training through reinforcement learning. This version emphasizes reinforcement learning as a primary training component, utilizing new tasks in math, competitive programming, science, tool use, and software engineering. The model trains in real environments for agentic skills, such as software engineering, where it operates within real repositories and is rewarded for passing tests.
Mellum2.1 Thinking demonstrates superior performance in various benchmarks compared to its predecessor and other models. It achieves high scores in coding benchmarks like LiveCodeBench v6 and HumanEval+, as well as in math benchmarks such as AIME 25/26 and GSM-Plus. The model also excels in agentic benchmarks, including SWE-bench Verified and Terminal-Bench 2.1, showcasing its enhanced capability in exploring codebases and editing files.
The model is released under the Apache 2.0 license, ensuring open-source availability. It supports serving with vLLM and is compatible with GGUF builds for llama.cpp, Ollama, and LM Studio. Mellum2.1 Thinking is ideal for developers and researchers seeking a robust tool for complex problem-solving and agentic tasks.
- CategoryAI Models & LLMs
- Pricing ModelOpen source
- DeploymentCloud
- PlatformsWeb
Mellum2.1 Thinking Highlights
12B Parameters
2.5B Active Parameters
28 Layers
2304 Hidden Size
98,304 Vocabulary Size
Reinforcement Learning
Open Source
Apache 2.0 License
Getting Started with Mellum2.1 Thinking
Access model page: Visit Hugging Face
Authenticate: Log in to your account
Configure: Set up model parameters
Send prompts: Input your queries
Fine-tune: Customize for specific tasks
Mellum2.1 Thinking's Use Cases
- Complex Coding Tasks
- Mathematical Reasoning
- Agentic Tasks
- Software Engineering
- Tool Use
