What is Mellum2.1 Thinking?
Mellum2.1 Thinking is JetBrains' newest open model, announced on 8 October 2026 as the next version of the Mellum2 model it open-sourced in June. It is a mixture-of-experts model with 12 billion parameters, 2.5 billion active per token, and a 131,072-token context. The architecture is unchanged from Mellum2; JetBrains spent the release on reinforcement learning, with millions of sandboxed runs inside real repositories where the model was rewarded when tests passed. It is text only and always reasons before it answers.
The weights are under Apache 2.0. JetBrains publishes BF16 weights, a multi-token prediction head for faster decoding in vLLM, and GGUF builds for llama.cpp, Ollama and LM Studio from 7.03 GB (MXFP4) and 8.07 GB (Q4_K_M) to 24.31 GB (BF16). On Hugging Face the main repository has 193 likes and 11,306 downloads in the last 30 days, and the GGUF repository 457,537.
JetBrains offers no hosted API for it; it is meant to be run on your own hardware.
How much does Mellum2.1 Thinking cost?
| Plan | Price | What you get |
|---|---|---|
| Weights | Free | Apache 2.0 |
| GGUF builds | Free | MXFP4, Q4_K_M, Q6_K, Q8_0 and BF16 |
Free weights under Apache 2.0. No hosted API; GGUF builds from 7.03 GB run locally.
Is Mellum2.1 Thinking worth it?
Mellum2.1 Thinking turns a model that could not work in a repository into one that can. On JetBrains' harness its SWE-bench Verified score went from 2.0% to 47.0%, Terminal-Bench 2.1 from 0.6% to 17.4% and SWE-bench Pro from 0.0% to 28.0%. LiveCodeBench v6 rose to 82.0, ahead of Qwen3.5-9B's 75.4 and Gemma 4 E4B's 69.4.
The gains are not enough to lead its class on agent work. In the same table Qwen3.5-9B scores 50.0 on SWE-bench Verified, 21.7 on Terminal-Bench 2.1 and 38.0 on SWE-bench Pro, and much more on GPQA Diamond, 77.8 against 64.6. All figures are JetBrains' own, run with the same open-source harness for every model.
Where it wins is speed. Only 2.5 billion parameters are active per token, so JetBrains says it serves almost twice as many tokens as Qwen3.5-9B under heavy load, and the multi-token prediction head makes single requests about 1.6 times faster. The Q4_K_M build is 8.07 GB and fits on a 12 GB graphics card.
The catch: it is a sub-agent, not a lead. A 17.4% Terminal-Bench score means most long terminal tasks fail, the context stops at 131,072 tokens, and it always thinks, which costs time on simple questions.
Our take: Mellum2.1 is the model to run many of at once, finding the failing test, drafting the fix, checking it, while a larger model plans the work. For a single local coding assistant, Qwen3.5-9B is still slightly smarter.
Where it earns its place
An engineer building a multi-agent coding pipeline
Run many cheap workers that find root causes and draft fixes
Serve Mellum2.1 with vLLM and the MTP head, and hand it narrow tasks such as tracing a failing test, leaving planning to a larger model. JetBrains says it serves almost twice the tokens of Qwen3.5-9B under load.
A developer at a company that bans cloud coding assistants
Keep a reasoning coding model entirely on a laptop or workstation
ollama run hf.co/JetBrains/Mellum2.1-12B-A2.5B-Thinking-GGUF:Q4_K_M pulls the 8.07 GB build, which fits a 12 GB GPU. Code never leaves the machine and the Apache 2.0 licence allows commercial use.
Before you commit
- Setup
- Download from Hugging Face and serve with vLLM, or pull a GGUF build into llama.cpp, Ollama or LM Studio; Q4_K_M is the recommended build.
- Integrations
- OpenAI-compatible endpoint through vLLM or Ollama; tool calling; JetBrains evaluated it in the Pi open-source agent harness with shell and file tools.
- Data and privacy
- No hosted service, so everything runs on hardware you control. Training used open RL datasets plus JetBrains-built tasks.
- Cost at scale
- Free weights; the 24.31 GB BF16 file fits one data-centre GPU, and the 8.07 GB Q4_K_M build runs on a 12 GB consumer card.
- Limits
- 131,072-token context, English only on the model card, and always-on reasoning. JetBrains reports 8.5% harmful answers on HarmBench, down from 21.5% for Mellum2.
What’s good
- SWE-bench Verified up from 2.0% to 47.0%, per JetBrains
- Only 2.5B active parameters, with an MTP head about 1.6 times faster per request
- Apache 2.0 licence
- Official GGUF builds from 7.03 GB
What’s not
- Qwen3.5-9B scores higher on SWE-bench, Terminal-Bench and GPQA in JetBrains' own table
- 17.4% on Terminal-Bench 2.1
- 131,072-token context
- No hosted API
Best for
Teams building coding agents that need fast, cheap sub-agents on their own hardware.
Mellum2.1 Thinking alternatives
Mellum2.1 Thinking: common questions
- Is Mellum2.1 Thinking worth it?
- A fast little coding agent for your own GPU that fixes routine bugs, though Qwen3.5-9B still edges it on the hard ones. We rate it 3.8 out of 5.
- How much does Mellum2.1 Thinking cost?
- Free weights under Apache 2.0. No hosted API; GGUF builds from 7.03 GB run locally.
- What are the downsides of Mellum2.1 Thinking?
- Qwen3.5-9B scores higher on SWE-bench, Terminal-Bench and GPQA in JetBrains' own table; 17.4% on Terminal-Bench 2.1; 131,072-token context; no hosted API.
- Who is Mellum2.1 Thinking best for?
- Teams building coding agents that need fast, cheap sub-agents on their own hardware.
- How much does Mellum2.1 Thinking cost for a team or in production?
- Free weights; the 24.31 GB BF16 file fits one data-centre GPU, and the 8.07 GB Q4_K_M build runs on a 12 GB consumer card.
- Is Mellum2.1 Thinking safe for company data?
- No hosted service, so everything runs on hardware you control. Training used open RL datasets plus JetBrains-built tasks.
- What are the best Mellum2.1 Thinking alternatives?
- The closest alternatives we have reviewed are Devstral Small 2, Qwen3.8-27B, gpt-oss-20b and Laguna S 2.1.
Answers checked Nov 7, 2026.