Skip to content

Mellum2.1 12B A2.5B Thinking VRAM requirements

How much VRAM Mellum2.1 12B A2.5B Thinking needs
4-bit weights: 6.9 GBWith 8K context: 7.7 GBWith 32K context: 8.0 GB8-bit, 8K context: 12.9 GB
Parameters
12.1B
Active per token
2.5B
Longest context
128K
Released
Sep 2026
License
Apache 2.0
Strongest at
Not rated

How much VRAM Mellum2.1 12B A2.5B Thinking needs

Memory Mellum2.1 12B A2.5B Thinking needs at each precision and context length
PrecisionWeightsWith 8KWith 32KWith 128K
16-bit22.6 GB23.5 GB23.8 GB25.1 GB
8-bit12.0 GB12.9 GB13.2 GB14.5 GB
6-bit9.3 GB10.1 GB10.5 GB11.8 GB
5-bit8.0 GB8.9 GB9.2 GB10.5 GB
4-bit6.9 GB7.7 GB8.0 GB9.4 GB
3-bit5.5 GB6.4 GB6.7 GB8.0 GB
2-bit4.2 GB5.1 GB5.4 GB6.7 GB
1.6-bit2.3 GB3.2 GB3.5 GB4.8 GB

One conversation, 16-bit context cache, runtime overhead included.

GPUs and Macs that run Mellum2.1 12B A2.5B Thinking

The best way Mellum2.1 12B A2.5B Thinking runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 40603-bit96K123 tok/s1.0 sPDF: Mellum2.1 12B A2.5B Thinking on RTX 4060 report
RTX 50603-bit96K179 tok/s0.8 sPDF: Mellum2.1 12B A2.5B Thinking on RTX 5060 report
RTX 30703-bit96K179 tok/s0.8 sPDF: Mellum2.1 12B A2.5B Thinking on RTX 3070 report
RTX 3080 10 GB5-bit48K199 tok/s0.5 sPDF: Mellum2.1 12B A2.5B Thinking on RTX 3080 10 GB report
RTX 3060 12 GB6-bit96K102 tok/s1.2 sPDF: Mellum2.1 12B A2.5B Thinking on RTX 3060 12 GB report
Arc B5806-bit96K85 tok/s0.8 sPDF: Mellum2.1 12B A2.5B Thinking on Arc B580 report
RTX 40706-bit96K133 tok/s0.5 sPDF: Mellum2.1 12B A2.5B Thinking on RTX 4070 report
RTX 4070 Super6-bit96K133 tok/s0.4 sPDF: Mellum2.1 12B A2.5B Thinking on RTX 4070 Super report
RTX 50706-bit96K166 tok/s0.5 sPDF: Mellum2.1 12B A2.5B Thinking on RTX 5070 report
RTX 4060 Ti 16 GB8-bit128K67 tok/s0.7 sPDF: Mellum2.1 12B A2.5B Thinking on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB8-bit128K99 tok/s0.7 sPDF: Mellum2.1 12B A2.5B Thinking on RTX 5060 Ti 16 GB report
Arc A770 16 GB8-bit128K81 tok/s0.8 sPDF: Mellum2.1 12B A2.5B Thinking on Arc A770 16 GB report
Radeon RX 9070 XT8-bit128K106 tok/s0.4 sPDF: Mellum2.1 12B A2.5B Thinking on Radeon RX 9070 XT report
RTX 4070 Ti Super8-bit128K136 tok/s0.4 sPDF: Mellum2.1 12B A2.5B Thinking on RTX 4070 Ti Super report
RTX 4080 Super8-bit128K146 tok/s0.3 sPDF: Mellum2.1 12B A2.5B Thinking on RTX 4080 Super report
RTX 5070 Ti8-bit128K169 tok/s0.4 sPDF: Mellum2.1 12B A2.5B Thinking on RTX 5070 Ti report
RTX 50808-bit128K178 tok/s0.3 sPDF: Mellum2.1 12B A2.5B Thinking on RTX 5080 report
Radeon RX 7900 XTX8-bit128K140 tok/s0.7 sPDF: Mellum2.1 12B A2.5B Thinking on Radeon RX 7900 XTX report
RTX 30908-bit128K174 tok/s0.4 sPDF: Mellum2.1 12B A2.5B Thinking on RTX 3090 report
RTX 40908-bit128K184 tok/s0.2 sPDF: Mellum2.1 12B A2.5B Thinking on RTX 4090 report
Radeon AI PRO R97008-bit128K106 tok/s0.4 sPDF: Mellum2.1 12B A2.5B Thinking on Radeon AI PRO R9700 report
RTX 50908-bit128K263 tok/s0.1 sPDF: Mellum2.1 12B A2.5B Thinking on RTX 5090 report
2× RTX 30908-bit128K174 tok/s0.3 sPDF: Mellum2.1 12B A2.5B Thinking on 2× RTX 3090 report
2× RTX 40908-bit128K184 tok/s0.1 sPDF: Mellum2.1 12B A2.5B Thinking on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit128K222 tok/s0.2 sPDF: Mellum2.1 12B A2.5B Thinking on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit128K178 tok/s0.2 sPDF: Mellum2.1 12B A2.5B Thinking on RTX 6000 Ada report
L40S8-bit128K165 tok/s0.3 sPDF: Mellum2.1 12B A2.5B Thinking on L40S report
2× RTX 50908-bit128K263 tok/s0.1 sPDF: Mellum2.1 12B A2.5B Thinking on 2× RTX 5090 report
4× RTX 30908-bit128K174 tok/s0.1 sPDF: Mellum2.1 12B A2.5B Thinking on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit128K263 tok/s0.1 sPDF: Mellum2.1 12B A2.5B Thinking on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB6-bit20K40 tok/s10.0 sPDF: Mellum2.1 12B A2.5B Thinking on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit128K32 tok/s10.0 sPDF: Mellum2.1 12B A2.5B Thinking on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit128K32 tok/s10.0 sPDF: Mellum2.1 12B A2.5B Thinking on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit128K63 tok/s5.0 sPDF: Mellum2.1 12B A2.5B Thinking on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit128K63 tok/s5.0 sPDF: Mellum2.1 12B A2.5B Thinking on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit128K63 tok/s5.0 sPDF: Mellum2.1 12B A2.5B Thinking on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit128K96 tok/s2.5 sPDF: Mellum2.1 12B A2.5B Thinking on M4 Max 128 GB report
M5 Max 128 GB8-bit128K103 tok/s0.7 sPDF: Mellum2.1 12B A2.5B Thinking on M5 Max 128 GB report
M3 Ultra 96 GB8-bit128K112 tok/s1.4 sPDF: Mellum2.1 12B A2.5B Thinking on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit128K112 tok/s1.4 sPDF: Mellum2.1 12B A2.5B Thinking on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit128K112 tok/s1.4 sPDF: Mellum2.1 12B A2.5B Thinking on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit128K137 tok/s0.4 sPDF: Mellum2.1 12B A2.5B Thinking on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit128K50 tok/s1.4 sPDF: Mellum2.1 12B A2.5B Thinking on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit128K50 tok/s1.4 sPDF: Mellum2.1 12B A2.5B Thinking on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit128K70 tok/s0.5 sPDF: Mellum2.1 12B A2.5B Thinking on RTX Spark report
DGX Spark8-bit128K64 tok/s0.5 sPDF: Mellum2.1 12B A2.5B Thinking on DGX Spark report
Dual DGX Spark8-bit128K64 tok/s0.3 sPDF: Mellum2.1 12B A2.5B Thinking on Dual DGX Spark report
DGX Station GB3008-bit128K452 tok/s0.1 sPDF: Mellum2.1 12B A2.5B Thinking on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit128K263 tok/s0.1 sPDF: Mellum2.1 12B A2.5B Thinking on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit128K263 tok/s0.1 sPDF: Mellum2.1 12B A2.5B Thinking on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit128K263 tok/s0.1 sPDF: Mellum2.1 12B A2.5B Thinking on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit128K282 tok/s0.3 sPDF: Mellum2.1 12B A2.5B Thinking on A100 80 GB SXM report
H100 SXM8-bit128K356 tok/s0.1 sPDF: Mellum2.1 12B A2.5B Thinking on H100 SXM report
4× H200 NVL8-bit128K405 tok/s0.1 sPDF: Mellum2.1 12B A2.5B Thinking on 4× H200 NVL report
8× H200 NVL8-bit128K405 tok/s0.1 sPDF: Mellum2.1 12B A2.5B Thinking on 8× H200 NVL report

Best fit is the highest quality, up to 8-bit, that leaves 16K of context. First reply is the wait for a 2K prompt. Speeds are rough, plus or minus 25%.

How to run it

llama.cpp
llama-server -hf JetBrains/Mellum2.1-12B-A2.5B-Thinking-GGUF
vLLM
vllm serve JetBrains/Mellum2.1-12B-A2.5B-Thinking
MLX (Mac)
mlx_lm.chat --model mlx-community/Mellum2.1-12B-A2.5B-Thinking-6bit

Questions

How much VRAM does Mellum2.1 12B A2.5B Thinking need?

Mellum2.1 12B A2.5B Thinking needs 6.9 GB for its weights at 4-bit, 7.7 GB with an 8K context and 8.0 GB with 32K. At 8-bit, 12.9 GB.

What is the minimum VRAM for Mellum2.1 12B A2.5B Thinking?

About 6.4 GB at 3-bit with an 8K context, the lowest precision we recommend for a model this size. The smallest machine here that runs it without spilling is the RTX 4060.

How big is Mellum2.1 12B A2.5B Thinking?

12.1B parameters (2.5B active per token): 22.6 GB of weights at 16-bit, 6.9 GB at 4-bit.

Can Mellum2.1 12B A2.5B Thinking run on an RTX 4090?

Yes. At 8-bit it holds 128K of context and writes about 184 tokens a second.

More models