What hardware do you need to run Mellum2.1 12B-A2.5B Thinking?
Mellum2.1 12B-A2.5B Thinking at Q4_K_M is 8.1 GB of weights, with a context ceiling of 128k tokens. JetBrains built this one for coding agents, and it asks less of a machine than anything else here aimed at that work: 12.15B parameters in the file, 2.5B of them read for each token, so the memory holds a 12B model and the speed is closer to a 2.5B one. The four-bit file loads in stock llama.cpp, Ollama and LM Studio with no patch. It writes its working out between <think> tags before it answers, and JetBrains's own example calls allow 81,920 tokens for the thinking and the reply together, which is output the machine generates itself and an API would bill for. On JetBrains's own table, every model run in thinking mode through the same harness, it leads Qwen3.5 9B on the standalone coding tests, 82.0 against 75.4 on LiveCodeBench v6, and trails it on all three agentic ones, 47.0 against 50.0 on SWE-bench Verified and 28.0 against 38.0 on SWE-bench Pro; on general knowledge it is well behind, 64.6 against 77.8 on GPQA Diamond. Those are the publisher's figures, self-reported, and nobody has placed this model on the index the rest of this site is ranked by. It is English-only and reads no images. The repository ships the model's multi-token prediction head as a separate 1.8 GB file for speculative decoding in llama.cpp, and the publisher's own benchmark gave 1.05x the decoding speed at full precision and no gain at all on a four-bit model with a quantised head, so it is not memory this page counts. On the cheapest machine that runs it, a Mac mini M6, 16GB, it generates 21 tok/s at 32k of context, worked out from memory bandwidth.
How good is Mellum2.1 12B-A2.5B Thinking, really?
It has not been placed on the intelligence index yet. See the whole table.
Head to head with a model that needs much the same memory at 32k of context: vs Gemma 4 12B.
What Mellum2.1 12B-A2.5B Thinking costs either way
Nobody rents Mellum2.1 12B-A2.5B Thinking by the token. The closest hosted match, Qwen3.5 9B, costs $0.08 per million input tokens and $0.12 per million output (OpenRouter, cheapest active endpoint, checked 2026-11-21). Buying a machine only beats that if you use it hard enough, for long enough, that the hardware price divides down below the rental bill. What a million tokens costs each way puts the two prices side by side.
Machines that run Mellum2.1 12B-A2.5B Thinking
Every machine here holds it to 128k, this model's own ceiling, so the choice between them is speed and price rather than how much you can keep in the window.
| Machine | Price | Speed at 32k | Longest context | Pay-back | |
|---|---|---|---|---|---|
| Mac mini M6, 16GB | $899 | 21 tok/s estimated | 128k | Pays back in 67 years | Run the numbers |
| Strix Halo Framework Desktop, 32GB | $1,269 | 65 tok/s estimated | 128k | Pays back in 91 years | Run the numbers |
| MacBook Air M5 (13-inch), 16GB | $1,299 | 21 tok/s estimated | 128k | Pays back in 97 years | Run the numbers |
| AMD Radeon AI PRO R9700, 32GB | $1,299card only | 153 tok/s estimated | 128k | Pays back in 93 years | Run the numbers |
| MacBook Pro M5 (14-inch), 16GB | $1,999 | 21 tok/s estimated | 128k | Pays back in 149 years | Run the numbers |
| NVIDIA GeForce RTX 5090, 32GB | $1,999card only | 272 tok/s estimated | 128k | Pays back in 144 years | Run the numbers |
| Mac Studio M5 Max, 36GB | $2,499 | 63 tok/s estimated | 128k | Pays back in 181 years | Run the numbers |
| DGX Spark GB10 Grace Blackwell, 128GB | $4,699 | 82 tok/s estimated | 128k | Pays back in 334 years | Run the numbers |
All 56 machines on this site run it at 32k, not just the eight in the table. One table sets every machine against the others: price, the memory its GPU can use, what it holds and how long it takes to pay for itself.
One machine per family, cheapest first. Speeds are measured where a public benchmark exists and estimated from memory bandwidth otherwise; the calculator says which for any configuration. The longest context is the longest setting the calculator offers that the machine still holds this model at, cache included, and no machine is shown taking it past its own 128k limit. Each one opens the calculator on that machine at that length.
For scale, the calculator starts from 80 tok/s for a hosted API and times a local machine against it. 3 of the 8 above reach it, and the slowest is 21 tok/s.
The cheapest machine that runs it is sold with 16 GB. There is a page on what 16 GB runs, machine by machine: it names the 6 machines here sold at that size, and the 19 current models the roomiest of them holds at 32k, this one among them.
It fits in less memory than that: the smallest size here that runs it at 32k is 12 GB, on the GeForce RTX 3060, 12GB at $329, card only, which hands a model 11 GB and is no longer sold.
Run Mellum2.1 12B-A2.5B Thinking at 128k on the Mac mini M6, 16GB
How much memory does Mellum2.1 12B-A2.5B Thinking need at 128k?
Every figure above is priced at 32k of context. At 128k, this model's own ceiling, Mellum2.1 12B-A2.5B Thinking needs 10 GB rather than 8.6 GB. The weights are the same either way; the key-value cache goes from 0.5 GB to 1.9 GB.
The window costs nothing in hardware: all 37 machines here still sold at a published price hold Mellum2.1 12B-A2.5B Thinking at both lengths, and the cheapest of them is the Mac mini M6, 16GB at $899.
On the Mac mini M6, 16GB it runs 21 tok/s estimated at 32k and 13 tok/s estimated at 128k. Every token generated reads the whole cache back, so a longer window is slower as well as bigger.
The specifics
- Parameters
- 12.2B, of which 2.5B are active per token
- Quantisation
- Q4_K_M
- Weights on disk
- 8.1 GB
- KV cache
- 0.5 GB at 32k context — Seven of the 28 layers grow with the context; the other 21 stop at a 1,024-token window, three windowed layers to every growing one. Both kinds cache the same geometry, four KV heads by 128 wide, so a token costs 2,048 bytes in whichever layer holds it. How weights and cache add up.
- Maximum context
- 128k tokens
- Licence
- Apache 2.0
- Sources
- JetBrains/Mellum2.1-12B-A2.5B-Thinking, JetBrains/Mellum2.1-12B-A2.5B-Thinking (config), JetBrains/Mellum2.1-12B-A2.5B-Thinking-GGUF, openrouter.ai