readygpu

MiniCPM5 2B

2.5B parameters · 128k context · Apache License 2.0 · released 2026-09-06

chatcodingagenticreasoning

In short

MiniCPM5 2B is an open model from openbmb with 2.5B parameters. At mlx-4bit, MiniCPM5 2B needs about 2 GB of memory at 8k context and 5 GB at 64k, counting the weights, the KV cache and runtime buffers. The cheapest hardware in the US that runs MiniCPM5 2B fully in memory at 64k context is the NVIDIA Tesla P40, $345 at Amazon, at about 39 tok/s (estimated).

Updated . Speeds are estimates; how they are worked out.

What to run it on in the US

Fully in memory at 64k context, priced at retailers today.

All hardware that runs it

Everything we price, best value first. Two cards split the model by layers.

HardwareQuantFitsDecodePricetok/s per 1,000
NVIDIA Tesla P40Q8_0fits39 tok/s$345113.0
Intel Arc B580Q8_0fits42 tok/s$346119.9
NVIDIA GeForce RTX 3060 12GBQ8_0fits41 tok/s$360112.8
Intel Arc A770 16GBQ8_0fits51 tok/s$400127.3
NVIDIA GeForce RTX 3080 10GBQ8_0fits86 tok/s$469182.6
AMD Radeon RX 7600 XTQ8_0fits29 tok/s$48060.0
AMD Radeon RX 9060 XT 16GBQ8_0fits32 tok/s$52061.5
AMD Radeon RX 7800 XTQ8_0fits62 tok/s$550113.5
NVIDIA GeForce RTX 5060 Ti 16GBQ8_0fits51 tok/s$61881.7
Intel Arc Pro B60Q8_0fits42 tok/s$65063.8
NVIDIA GeForce RTX 4070 SuperQ8_0fits57 tok/s$69981.2
Mac mini M4 16GBQ8_0fits14 tok/s$74919.0
NVIDIA GeForce RTX 5070Q8_0fits76 tok/s$750101.1
NVIDIA GeForce RTX 4060 Ti 16GBQ8_0fits33 tok/s$77042.2
AMD Radeon RX 7900 XTQ8_0fits80 tok/s$800100.0
AMD Radeon RX 9070 XTQ8_0fits64 tok/s$81578.5
NVIDIA GeForce RTX 4070 Ti SuperQ8_0fits76 tok/s$98477.1
NVIDIA GeForce RTX 5070 TiQ8_0fits101 tok/s$995101.5
NVIDIA GeForce RTX 4080 SuperQ8_0fits83 tok/s$1,19969.2
Mac mini M4 32GBQ8_0fits14 tok/s$1,42010.0
AMD Radeon RX 7900 XTXQ8_0fits96 tok/s$1,53362.6
NVIDIA GeForce RTX 3090Q8_0fits106 tok/s$1,60065.9
NVIDIA GeForce RTX 5080Q8_0fits108 tok/s$1,60067.7
AMD Radeon AI PRO R9700Q8_0fits64 tok/s$1,70037.6
NVIDIA GeForce RTX 3090 TiQ8_0fits114 tok/s$2,06055.2
NVIDIA GeForce RTX 4090Q8_0fits114 tok/s$2,07354.8
Mac mini M4 Pro 48GBQ8_0fits32 tok/s$2,14915.0
Mac mini M4 Pro 64GBQ8_0fits32 tok/s$2,90011.1
Mac Studio M4 Max 64GBQ8_0fits65 tok/s$3,40019.0
Framework Desktop Ryzen AI Max+ 395 128GBQ8_0fits26 tok/s$3,4497.4
GMKtec EVO-X2 Ryzen AI Max+ 395 128GBQ8_0fits26 tok/s$3,6507.0
Mac Studio M3 Ultra 96GBQ8_0fits97 tok/s$4,39922.0
NVIDIA GeForce RTX 5090Q8_0fits202 tok/s$4,92141.1
Mac Studio M4 Max 128GBQ8_0fits65 tok/s$4,99912.9
ASUS Ascent GX10 128GBQ8_0fits31 tok/s$6,1005.0
NVIDIA DGX Spark 128GBQ8_0fits31 tok/s$7,3994.2
NVIDIA RTX PRO 6000 BlackwellQ8_0fits202 tok/s$7,99725.3
Mac Studio M3 Ultra 256GBQ8_0fits97 tok/s$8,49911.4
Mac Studio M3 Ultra 512GBQ8_0fits97 tok/s$23,0004.2
Framework Desktop Ryzen AI Max+ 395 64GBQ8_0fits26 tok/s––

Memory needed

Weights, KV cache and runtime buffers at each context length (KV cache at fp16).

QuantFile@8k@32k@64k@128k
mlx-4bitmlx1.3 GB2.4 GB3.5 GB5.0 GB8.0 GB
Q4_K_Mgguf1.5 GB2.6 GB3.7 GB5.2 GB8.2 GB
Q5_K_Mgguf1.8 GB2.9 GB4.0 GB5.5 GB8.5 GB
Q6_Kgguf2.0 GB3.0 GB4.2 GB5.7 GB8.7 GB
mlx-8bitmlx2.5 GB3.6 GB4.7 GB6.2 GB9.2 GB
Q8_0gguf2.5 GB3.6 GB4.7 GB6.2 GB9.2 GB

Run it

Copy a command for the runtime you use.

  • mlx-4bit
    • LM Studiolms get mlx-community/MiniCPM5-2B-mlx-4Bit
    • MLXmlx_lm.server --model mlx-community/MiniCPM5-2B-mlx-4Bit
  • Q4_K_M
    • Ollamaollama run hf.co/bartowski/MiniCPM5-2B-GGUF:Q4_K_M
    • LM Studiolms get bartowski/MiniCPM5-2B-GGUF@q4_k_m
    • llama.cppllama-server -hf bartowski/MiniCPM5-2B-GGUF:Q4_K_M
  • Q5_K_M
    • Ollamaollama run hf.co/bartowski/MiniCPM5-2B-GGUF:Q5_K_M
    • LM Studiolms get bartowski/MiniCPM5-2B-GGUF@q5_k_m
    • llama.cppllama-server -hf bartowski/MiniCPM5-2B-GGUF:Q5_K_M
  • Q6_K
    • Ollamaollama run hf.co/bartowski/MiniCPM5-2B-GGUF:Q6_K
    • LM Studiolms get bartowski/MiniCPM5-2B-GGUF@q6_k
    • llama.cppllama-server -hf bartowski/MiniCPM5-2B-GGUF:Q6_K
  • mlx-8bit
    • LM Studiolms get mlx-community/MiniCPM5-2B-8bit
    • MLXmlx_lm.server --model mlx-community/MiniCPM5-2B-8bit
  • Q8_0
    • Ollamaollama run hf.co/bartowski/MiniCPM5-2B-GGUF:Q8_0
    • LM Studiolms get bartowski/MiniCPM5-2B-GGUF@q8_0
    • llama.cppllama-server -hf bartowski/MiniCPM5-2B-GGUF:Q8_0

Notes

Dense dual-language (en/zh) 2B model with standard LlamaForCausalLM architecture, aimed at on-device use. This is the final RL+OPD post-trained BF16 release; GGUF, MLX and GPTQ variants are separate repos.

Questions

How much VRAM does MiniCPM5 2B need?
At mlx-4bit, MiniCPM5 2B needs about 2 GB of memory at 8k context and 5 GB at 64k, counting the weights, the KV cache and runtime buffers.
What is the best GPU for running MiniCPM5 2B locally?
The cheapest hardware in the US that runs MiniCPM5 2B fully in memory at 64k context is the NVIDIA Tesla P40, $345 at Amazon, at about 39 tok/s (estimated). The best value is the NVIDIA GeForce RTX 3080 10GB (about 86 tok/s for $469). The fastest we price is the NVIDIA GeForce RTX 5090, at about 202 tok/s.
Can I run MiniCPM5 2B 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 108 tok/s (estimated).
Can I run MiniCPM5 2B 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 114 tok/s (estimated).
How do I run MiniCPM5 2B locally?
With Ollama: ollama run hf.co/bartowski/MiniCPM5-2B-GGUF:Q4_K_M (the Q4_K_M build). LM Studio and llama.cpp commands are on this page.
What licence is MiniCPM5 2B released under?
Apache License 2.0, which allows commercial use.