Mellum2 12B A2.5B Instruct SFT — Hardware Requirements & GPU Compatibility
Chat
Mellum2 12B A2.5B Instruct SFT is a 12.1B-parameter open language model from JetBrains in the Mellum family. It supports a context window of up to 131,072 tokens. At Q4_K_M it needs about 7.66 GB of VRAM — see which GPUs and Macs can run it below.
Runs great on a 24 GB GPU
Runs from 8 GB (Q4_K_M · 32K)Best on 24 GB (Q4_K_M · ~257 tok/s · 128K)Macs from 16 GB (Q4_K_M · ~60 tok/s)
Apache 2.0
Reference GPU · RTX 4090 24 GB
Estimates, ±20% · how we estimate
Run Mellum2 12B A2.5B Instruct SFT
One recommended way. Every other method, OS and quant is in the step-by-step guide.
Recommended
vllm serve JetBrains/Mellum2-12B-A2.5B-Instruct-SFT --max-model-len 16384vLLM repo with its own quantisation
Download: Weights on Hugging Face
Other ways
each opens step-by-stepRun in cloud
Fits on RTX 3060 12GB (4 GB headroom) · Q4_K_M
- Generation speed
- ~123 tok/s
- generation speed
- Cost per 1M output tokens
- $0.19
- per 1M output tokens
How Much VRAM Does Mellum2 12B A2.5B Instruct SFT Need?
Select a quantization level; the listed hardware adheres to it.
| Quantization | VRAM | +32K | On RTX 4090 (reference) | |
|---|---|---|---|---|
| 5.5 GB | 5.8 GB | 2-bit · K-quant — 2-bit quantization with K-quant improvements | Fits: 128K context · ~309 tok/s | |
| 6.3 GB | 6.5 GB | 3-bit medium — 3-bit medium quantization | Fits: 128K context · ~288 tok/s | |
| 7.7 GB | 7.9 GB | 4-bit medium — 4-bit medium quantization — most popular sweet spot | Fits: 128K context · ~257 tok/s | |
| 9.0 GB | 9.2 GB | 5-bit medium — 5-bit medium quantization — good quality/size tradeoff | Fits: 128K context · ~232 tok/s | |
| 10.4 GB | 10.6 GB | 6-bit, very good — 6-bit quantization, very good quality | Fits: 128K context · ~212 tok/s | |
| 12.5 GB | 12.7 GB | 8-bit, near-lossless — 8-bit quantization, near-lossless | Fits: 128K context · ~186 tok/s | |
| 24.7 GB | 24.9 GB | 16-bit · bfloat — Brain floating point 16 — preferred for training | Does not fit: needs 24.8 GB |
Selected Q4_K_M
4-bit medium quantization7.8GBat 16K context
- Weights
- 7.3 GB
- KV cache
- 0.2 GB
- Overhead
- 0.3 GB
Fits RTX 4090 (reference) · 16.3 GB free · ~257 tok/s
≥50 tok/s needs ≥128 GB/s · RTX 5090 ~350, RTX 3090 Ti ~257
Which Hardware Can Run Mellum2 12B A2.5B Instruct SFT?
Q4_K_M · 7.7 GBMellum2 12B A2.5B Instruct SFT (Q4_K_M) requires 7.7 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 10+ GB is recommended. Using the full 131K context window can add up to 1.0 GB, bringing total usage to 8.7 GB. 52 GPUs can run it, including NVIDIA GeForce RTX 5090, NVIDIA GeForce RTX 3090 Ti, NVIDIA GeForce RTX 3080. 151 devices with unified memory can run Mellum2 12B A2.5B Instruct SFT, including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB), NVIDIA Jetson Orin Nano 8GB (Super).
≤ 11 GB
≤ 11 GB tier →fits · 0.3 GB free · up to 32K ctx
16 GPUs · 1 device family
RTX 3080~215Simulate output on RTX 3080RTX 2080 Ti~186Simulate output on RTX 2080 TiRTX 3070 Ti~184Simulate output on RTX 3070 TiOrin Nano~40Simulate output on NVIDIA Jetson Orin Nano 8GB (Super)20 GB
fits · 12.3 GB free · up to 128K ctx
1 GPU
RX 7900 XT~114Simulate output on RX 7900 XT32–40 GB
fits · 24.3 GB free · up to 128K ctx
1 GPU · 3 Mac classes · 4 device families
RTX 5090~350Simulate output on RTX 5090M5 Max36 GB~96Simulate output on Mac Studio M5 Max (36 GB)M532 GB~50Simulate output on MacBook Air 13" M5 (32 GB)Orin~76Simulate output on NVIDIA Jetson AGX Orin 32GBSnapdragon X Elite (X1E-80-100)~49Simulate output on Snapdragon X Elite Copilot+ PC
By VRAM tier · typical GPUs and Macs at Q4_K_M
Alternatives to Mellum2 12B A2.5B Instruct SFT
Mellum family · 8 models →Similar models
6| Model | Score | VRAM | tok/s | Why it is similar |
|---|---|---|---|---|
| Gemma 3 12B ITGoogle | 74.2 | 10.9 GB | ~82 | same size |
| Gemma 4 12BGoogle | — | 9 GB | ~80 | same size |
| Mistral Nemo Base 2407Mistral AI | — | 11 GB | ~81 | same size |
| Baichuan2 13B Basebaichuan-inc | — | 11.5 GB | ~76 | same size |
| Llama 2 13B Chat HFMeta | — | 11.5 GB | ~76 | same size |
| Vicuna 13B V1.3LMSYS | — | 8.6 GB | ~76 | same size |
Specifications
- Publisher
- JetBrains
- Family
- Mellum
- Parameters
- 12.1B
- Active parameters
- 2.4B per token
- Architecture
- MellumForCausalLM
- Context Length
- 131,072 tokens
- Vocabulary Size
- 98,304
- Release Date
- 2026-05-26
- License
- Apache 2.0
Frequently Asked Questions
How much VRAM does Mellum2 12B A2.5B Instruct SFT need?
Mellum2 12B A2.5B Instruct SFT requires 7.7 GB of VRAM at Q4_K_M, or 24.7 GB at BF16. Full 131K context adds up to 1.0 GB (8.7 GB total).
VRAM = Weights + KV Cache + Overhead
Weights = 12.1B × 4.8 bits ÷ 8 = 7.3 GB
KV Cache + Overhead ≈ 0.4 GB (at 2K context + ~0.3 GB framework)
Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.
KV Cache + Overhead ≈ 1.4 GB (at full 131K context)
VRAM usage by quantization
Can NVIDIA GeForce RTX 4090 run Mellum2 12B A2.5B Instruct SFT?
Yes, at Q8_0 (12.5 GB) or lower. Higher quantizations like BF16 (24.7 GB) exceed the NVIDIA GeForce RTX 4090's 24 GB.
What's the best quantization for Mellum2 12B A2.5B Instruct SFT?
For Mellum2 12B A2.5B Instruct SFT, Q4_K_M (7.7 GB) offers the best balance of quality and VRAM usage. Q5_K_M (9.0 GB) provides better quality if you have the VRAM. The smallest option is Q2_K at 5.5 GB.
VRAM requirement by quantization
★ Recommended — best balance of quality and VRAM usage.
Can I run Mellum2 12B A2.5B Instruct SFT on a Mac?
Yes — Mac Mini M2 Pro (16 GB) and 133 other Macs can run Mellum2 12B A2.5B Instruct SFT. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 7.7 GB of usable unified memory (RAM minus macOS overhead).
Can I run Mellum2 12B A2.5B Instruct SFT locally?
Yes — Mellum2 12B A2.5B Instruct SFT can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.
How fast is Mellum2 12B A2.5B Instruct SFT?
At Q4_K_M, Mellum2 12B A2.5B Instruct SFT can reach ~169 tok/s on AMD Instinct MI350X. On NVIDIA GeForce RTX 4090: ~257 tok/s. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.
tok/s = 1000 ÷ (active GB ÷ (bandwidth GB/s × efficiency) × 1000 + layers × routing ms)
Mixture-of-Experts: just the active experts get read for each token, along with a fixed routing cost per layer.
Example: NVIDIA B200 → 1.5 GB active ÷ (8000 × 0.65) = 0.296 ms, plus 28 layers × 0.055 ms = 1.540 ms, so 1000 ÷ 1.836 ms = ~545 tok/s
Estimated speed at Q4_K_M (7.7 GB)
Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.
What's the download size of Mellum2 12B A2.5B Instruct SFT?
At Q4_K_M, the download is about 7.29 GB. The full-precision BF16 version is 24.30 GB. The smallest option (Q2_K) is 5.16 GB.
Which GPUs can run Mellum2 12B A2.5B Instruct SFT?
52 consumer GPUs can run Mellum2 12B A2.5B Instruct SFT at Q4_K_M. Top options include AMD Radeon RX 6700 XT, AMD Radeon RX 6800, AMD Radeon RX 6800 XT, AMD Radeon RX 7600. 36 GPUs have plenty of headroom for comfortable inference.
Which devices can run Mellum2 12B A2.5B Instruct SFT?
151 devices with unified memory can run Mellum2 12B A2.5B Instruct SFT at Q4_K_M (7.7 GB), including AMD Ryzen AI 9 HX 370 (Strix Point) Laptop, ASUS Ascent GX10, Beelink GTR9 Pro (Ryzen AI Max+ 395, 128 GB), Framework Desktop (Ryzen AI Max+ 395, 128 GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.