Mellum2.1 12B A2.5B Thinking
12.1B total, 2.5B active · 128k context · Apache 2.0 · released 2026-09-20
chatcodingagenticreasoning
In short
Mellum2.1 12B A2.5B Thinking is an open model from JetBrains with 12.1B parameters (2.5B active per token). At Q6_K, Mellum2.1 12B A2.5B Thinking needs about 12 GB of memory at 8k context and 13 GB at 64k, counting the weights, the KV cache and runtime buffers. The cheapest hardware in the UK that runs Mellum2.1 12B A2.5B Thinking fully in memory at 64k context is the NVIDIA Tesla P40, £261 at eBay, at about 35 tok/s (estimated).
Updated . Speeds are estimates; how they are worked out.
What to run it on in the UK
Fully in memory at 64k context, priced at retailers today.
Cheapest
NVIDIA Tesla P4035tok/sest.Q8_0
£261
Best value
NVIDIA Tesla P4035tok/sest.Q8_0
£261
Fastest
NVIDIA GeForce RTX 5090183tok/sest.Q8_0
£5,000
Mac
Mac mini M4 Pro 48GB29tok/sest.Q8_0
£3,000
All hardware that runs it
Everything we price, best value first. Two cards split the model by layers.
| Hardware | Quant | Fits | Decode | Price | tok/s per 1,000 |
|---|---|---|---|---|---|
| NVIDIA Tesla P40 | Q8_0 | fits | 35 tok/s | £261 | 135.4 |
| AMD Radeon RX 7600 XT | Q8_0 | fits | 26 tok/s | £348 | 74.9 |
| AMD Radeon RX 9060 XT 16GB | Q8_0 | fits | 29 tok/s | £499 | 58.1 |
| NVIDIA GeForce RTX 5060 Ti 16GB | Q8_0 | fits | 46 tok/s | £550 | 83.1 |
| AMD Radeon RX 7800 XT | Q8_0 | fits | 57 tok/s | £568 | 99.5 |
| Intel Arc A770 16GB | Q8_0 | fits | 42 tok/s | £588 | 71.9 |
| AMD Radeon RX 9070 XT | Q8_0 | fits | 58 tok/s | £670 | 86.6 |
| Intel Arc Pro B60 | Q8_0 | fits | 34 tok/s | £742 | 46.4 |
| AMD Radeon RX 7900 XT | Q8_0 | fits | 73 tok/s | £789 | 91.9 |
| NVIDIA GeForce RTX 4070 Ti Super | Q8_0 | fits | 69 tok/s | £950 | 72.2 |
| NVIDIA GeForce RTX 5070 Ti | Q8_0 | fits | 92 tok/s | £975 | 93.8 |
| NVIDIA GeForce RTX 4060 Ti 16GB | Q8_0 | fits | 29 tok/s | £977 | 30.1 |
| AMD Radeon RX 7900 XTX | Q8_0 | fits | 87 tok/s | £999 | 87.1 |
| NVIDIA GeForce RTX 5080 | Q8_0 | fits | 98 tok/s | £1,199 | 81.7 |
| Mac mini M4 32GB | Q8_0 | fits | 13 tok/s | £1,200 | 10.7 |
| NVIDIA GeForce RTX 4080 Super | Q8_0 | fits | 75 tok/s | £1,429 | 52.6 |
| AMD Radeon AI PRO R9700 | Q8_0 | fits | 58 tok/s | £1,674 | 34.6 |
| NVIDIA GeForce RTX 3090 | Q8_0 | fits | 96 tok/s | £1,707 | 56.0 |
| NVIDIA GeForce RTX 3090 Ti | Q8_0 | fits | 103 tok/s | £2,199 | 46.8 |
| Mac mini M4 Pro 48GB | Q8_0 | fits | 29 tok/s | £3,000 | 9.7 |
| GMKtec EVO-X2 Ryzen AI Max+ 395 128GB | Q8_0 | fits | 23 tok/s | £3,320 | 7.0 |
| Mac Studio M4 Max 64GB | Q8_0 | fits | 58 tok/s | £3,399 | 17.2 |
| NVIDIA GeForce RTX 5090 | Q8_0 | fits | 183 tok/s | £5,000 | 36.6 |
| Mac Studio M3 Ultra 96GB | Q8_0 | fits | 88 tok/s | £5,499 | 15.9 |
| NVIDIA GeForce RTX 4090 | Q8_0 | fits | 103 tok/s | £6,015 | 17.1 |
| Mac Studio M3 Ultra 256GB | Q8_0 | fits | 88 tok/s | £8,499 | 10.3 |
| Mac Studio M3 Ultra 512GB | Q8_0 | fits | 88 tok/s | £17,495 | 5.0 |
| NVIDIA RTX PRO 6000 Blackwell | Q8_0 | fits | 183 tok/s | £29,504 | 6.2 |
| Mac Studio M4 Max 128GB | Q8_0 | fits | 58 tok/s | – | – |
| Mac mini M4 Pro 64GB | Q8_0 | fits | 29 tok/s | – | – |
| Framework Desktop Ryzen AI Max+ 395 128GB | Q8_0 | fits | 23 tok/s | – | – |
| NVIDIA DGX Spark 128GB | Q8_0 | fits | 28 tok/s | – | – |
| ASUS Ascent GX10 128GB | Q8_0 | fits | 28 tok/s | – | – |
| Framework Desktop Ryzen AI Max+ 395 64GB | Q8_0 | fits | 23 tok/s | – | – |
| Intel Arc B580 | Q6_K | offload | 26 tok/s | £349 | 74.9 |
| NVIDIA GeForce RTX 3060 12GB | Q6_K | offload | 30 tok/s | £390 | 76.7 |
| NVIDIA GeForce RTX 3080 10GB | Q6_K | offload | 27 tok/s | £466 | 57.1 |
| NVIDIA GeForce RTX 5070 | Q6_K | offload | 43 tok/s | £695 | 62.2 |
| NVIDIA GeForce RTX 4070 Super | Q6_K | offload | 37 tok/s | £874 | 42.2 |
1 other card or machine we price cannot hold it at this context.
Memory needed
Weights, KV cache and runtime buffers at each context length (KV cache at fp16).
| Quant | File | @8k | @32k | @64k | @128k |
|---|---|---|---|---|---|
| Q6_Kgguf | 10.7 GB | 11.6 GB | 12.0 GB | 12.7 GB | 13.9 GB |
| Q8_0gguf | 12.0 GB | 12.9 GB | 13.4 GB | 14.0 GB | 15.3 GB |
Run it
Copy a command for the runtime you use.
- Q6_K
- Ollama
ollama run hf.co/bartowski/Mellum2.1-12B-A2.5B-Thinking-GGUF:Q6_K - LM Studio
lms get bartowski/Mellum2.1-12B-A2.5B-Thinking-GGUF@q6_k - llama.cpp
llama-server -hf bartowski/Mellum2.1-12B-A2.5B-Thinking-GGUF:Q6_K
- Ollama
- Q8_0
- Ollama
ollama run hf.co/bartowski/Mellum2.1-12B-A2.5B-Thinking-GGUF:Q8_0 - LM Studio
lms get bartowski/Mellum2.1-12B-A2.5B-Thinking-GGUF@q8_0 - llama.cpp
llama-server -hf bartowski/Mellum2.1-12B-A2.5B-Thinking-GGUF:Q8_0
- Ollama
Notes
Thinking model; serve with vLLM using the qwen3 reasoning parser (and hermes tool-call parser for tools). GGUF builds and the MTP head are not yet released.
Questions
- How much VRAM does Mellum2.1 12B A2.5B Thinking need?
- At Q6_K, Mellum2.1 12B A2.5B Thinking needs about 12 GB of memory at 8k context and 13 GB at 64k, counting the weights, the KV cache and runtime buffers.
- What is the best GPU for running Mellum2.1 12B A2.5B Thinking locally?
- The cheapest hardware in the UK that runs Mellum2.1 12B A2.5B Thinking fully in memory at 64k context is the NVIDIA Tesla P40, £261 at eBay, at about 35 tok/s (estimated). The fastest we price is the NVIDIA GeForce RTX 5090, at about 183 tok/s.
- Can I run Mellum2.1 12B A2.5B Thinking on a 16 GB GPU?
- Yes: at Q8_0 it fits fully in 16 GB at 64k context. On the NVIDIA GeForce RTX 5080 it runs at about 98 tok/s (estimated).
- Can I run Mellum2.1 12B A2.5B Thinking on a 24 GB GPU?
- Yes: at Q8_0 it fits fully in 24 GB at 64k context. On the NVIDIA GeForce RTX 3090 Ti it runs at about 103 tok/s (estimated).
- How do I run Mellum2.1 12B A2.5B Thinking locally?
- With Ollama: ollama run hf.co/bartowski/Mellum2.1-12B-A2.5B-Thinking-GGUF:Q6_K (the Q6_K build). LM Studio and llama.cpp commands are on this page.
- What licence is Mellum2.1 12B A2.5B Thinking released under?
- Apache 2.0, which allows commercial use.