Model library / JetBrains
Mellum2.1 Thinking is a 12 billion-parameter mixture-of-experts language model designed for complex agentic coding, math and reasoning tasks.
Mixture of experts, count not published
Context length in tokens, as published
Weights at the released precision
Licence declared on the repository
Overview
Mellum2.1 Thinking is a 12 billion-parameter model with 2.5 billion active parameters, built by JetBrains. The publisher describes it as a 'thinking' model intended for hard non-agentic problems and for agentic tasks such as repository navigation, command execution and tool use, especially in software engineering contexts.
The model comprises 28 layers, a hidden size of 2 304 and a context window of 131 072 tokens, with a sliding-window mechanism on three-quarters of the layers. It uses bfloat16 precision, a vocabulary of 98 304 tokens, and is released under the Apache 2.0 licence, requiring appropriate hardware for the 12 billion-parameter size.
Specification
| Repository | JetBrains/Mellum2.1-12B-A2.5B-Thinking |
| Publisher | JetBrains |
| Published | 20 September 2026 |
| Architecture | MellumForCausalLM |
| Model type | mellum |
| Parameters | 12,149,923,072 |
| Layers | 28 |
| Hidden size | 2,304 |
| Attention heads | 32, 4 key/value heads |
| Context length | 131,072 tokens |
| Vocabulary | 98,304 tokens |
| Weight format | bfloat16 |
| Licence | apache-2.0 |
| Base model | JetBrains/Mellum2-12B-A2.5B-Base |
Memory
Parameter count multiplied by the bytes each format uses. The key-value cache comes on top.
| Precision | Weights | Notes |
|---|---|---|
| bf16 | 22.6 GiB | Full precision as released |
| fp8 | 11.3 GiB | 8-bit, near-lossless on most models |
| int4 | 5.7 GiB | 4-bit, smallest footprint |
The cache costs 56 KiB per token at 16-bit, so the full 131,072-token context of one request needs about 7.0 GiB. Concurrency multiplies that number, not the weights.
That figure is an upper bound: this model uses sliding-window attention with a 1,024-token window on some layers, so long requests cache less than this.
Hardware
At the released precision, with 90% of the card given to the server and 2.5 GiB kept for the runtime. The context column is what is left for the cache on that setup.
| GPU | Cards needed | Free for cache | Context it holds |
|---|---|---|---|
| RTX 4090 24GB | 2 | 15.6 GiB | the full 131,072 tokens |
| RTX 5090 32GB | 1 | 3.7 GiB | about 68,700 tokens |
| L40S 48GB | 1 | 18.1 GiB | the full 131,072 tokens |
| A100 80GB | 1 | 46.9 GiB | the full 131,072 tokens |
| H100 80GB | 1 | 46.9 GiB | the full 131,072 tokens |
| RTX PRO 6000 96GB | 1 | 61.3 GiB | the full 131,072 tokens |
| H200 141GB | 1 | 101.8 GiB | the full 131,072 tokens |
Serving
A vLLM launch line for the shape above. Check the model card for a runtime the publisher recommends.
vllm serve JetBrains/Mellum2.1-12B-A2.5B-Thinking \ --max-model-len 131072 \ --gpu-memory-utilization 0.90
Questions
The weights take about 23 GiB at the released precision. That fits on one RTX 5090 32GB, which leaves roughly 4 GiB for the key-value cache.
The repository declares apache-2.0. Read the licence text before commercial use: the name alone does not tell you what is allowed.
The configuration allows 131,072 tokens. Whether the whole window is usable depends on the memory left for the cache, which the hardware table works out per card.
This page is a reference, and Aiqre does not serve this model today. We deploy open-weight models on dedicated EU hardware on request, so if you want an OpenAI-compatible endpoint for it, email hello@aiqre.com with the model name and roughly what volume you expect.
Also in the library
Granite Switch 4.1 8B is a modular 8.8 billion-parameter language model that combines a base LLM, decision and safety adapters, a…
LFM2.5-24B-A2B is a 24 billion-parameter MoE reasoning model with 2.45 billion active parameters for on-device agentic tasks.
LFM2.5-24B-A2B-Base is a 24 billion-parameter mixture-of-experts model from LiquidAI aimed at on-device deployment and heavy fine…
ALoDLM-8B is an 8.2 billion-parameter diffusion language model from Amazon for research on adaptive computation, maths, and code …
Hosting
Tell us the model and roughly what volume you expect. We reply with a price and an OpenAI-compatible endpoint, with your prompts never stored.