readygpu

AREX 2

27.4B parameters · 256k context · Apache License 2.0 · released 2026-09-29

agenticreasoningcodingvisionchat

In short

AREX 2 is an open model from BAAI with 27.4B parameters. At Q4_K_M, AREX 2 needs about 17 GB of memory at 8k context and 19 GB at 32k, counting the weights, the KV cache and runtime buffers. The cheapest hardware in the US that runs AREX 2 fully in memory at 32k context is the AMD Radeon RX 7900 XT, $926 at Amazon, at about 23 tok/s (estimated).

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

Cheapest way to run it in the US today

AMD Radeon RX 7900 XT, used at eBay

Runs AREX 2 at Q4_K_M fully in memory at 32k context, about 23 tok/sest.

$849

In stock, checked

What to run it on in the US

Fully in memory at 32k 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 P40Q5_K_Mfits9 tok/s$37724.7
AMD Radeon RX 7900 XTQ4_K_Mfits23 tok/s$92624.4
Intel Arc Pro B60Q5_K_Mfits10 tok/s$1,0009.9
Mac mini M4 32GBQ4_K_Mfits4 tok/s$1,3303.0
AMD Radeon RX 7900 XTXQ5_K_Mfits23 tok/s$1,40016.4
NVIDIA RTX PRO 4000 BlackwellQ5_K_Mfits18 tok/s$1,48012.2
NVIDIA GeForce RTX 3090Q5_K_Mfits25 tok/s$1,77714.2
Intel Arc Pro B70Q8_0fits10 tok/s$1,8405.3
NVIDIA GeForce RTX 3090 TiQ5_K_Mfits27 tok/s$2,03313.4
NVIDIA GeForce RTX 4090Q5_K_Mfits27 tok/s$2,07313.1
Mac mini M4 Pro 48GBQ8_0fits6 tok/s$2,0992.7
AMD Radeon AI PRO R9700Q8_0fits11 tok/s$2,6994.2
GMKtec EVO-X2 Ryzen AI Max+ 395 128GBQ8_0fits5 tok/s$3,5001.3
Framework Desktop Ryzen AI Max+ 395 128GBQ8_0fits5 tok/s$3,5001.3
Beelink GTR9 Pro Ryzen AI Max+ 395 128GBQ8_0fits5 tok/s$3,6991.2
Mac mini M4 Pro 64GBQ8_0fits6 tok/s$3,8501.5
AMD Radeon PRO W7900Q8_0fits15 tok/s$4,0003.8
Mac Studio M4 Max 64GBQ8_0fits12 tok/s$4,0002.9
NVIDIA RTX PRO 4500 BlackwellQ8_0fits18 tok/s$4,3994.1
Mac Studio M3 Ultra 96GBQ8_0fits17 tok/s$4,5503.8
Mac Studio M4 Max 128GBQ8_0fits12 tok/s$5,4002.1
NVIDIA RTX A6000Q8_0fits15 tok/s$5,9892.6
ASUS Ascent GX10 128GBQ8_0fits6 tok/s$6,1000.9
NVIDIA GeForce RTX 5090Q8_0fits36 tok/s$6,9995.1
NVIDIA RTX PRO 5000 Blackwell 48GBQ8_0fits27 tok/s$7,2903.7
NVIDIA DGX Spark 128GBQ8_0fits6 tok/s$7,6000.7
NVIDIA RTX 6000 Ada GenerationQ8_0fits19 tok/s$8,4952.3
Mac Studio M3 Ultra 256GBQ8_0fits17 tok/s$9,5001.8
NVIDIA RTX PRO 5000 Blackwell 72GBQ8_0fits27 tok/s$9,5002.8
NVIDIA RTX PRO 6000 BlackwellQ8_0fits36 tok/s$17,6902.0
Mac Studio M3 Ultra 512GBQ8_0fits17 tok/s$27,5000.6
Framework Desktop Ryzen AI Max+ 395 64GBQ8_0fits5 tok/s––
Lenovo ThinkStation PGX GB10 128GBQ8_0fits6 tok/s––
NVIDIA GeForce RTX 3060 12GBQ4_K_Moffload5 tok/s$36813.3
NVIDIA GeForce RTX 5050Q4_K_Moffload4 tok/s$3909.5
NVIDIA GeForce RTX 4060Q4_K_Moffload4 tok/s$4228.5
Intel Arc B580Q4_K_Moffload5 tok/s$42511.3
Intel Arc A770 16GBQ4_K_Moffload8 tok/s$42517.6
NVIDIA GeForce RTX 5060Q4_K_Moffload4 tok/s$4828.1
NVIDIA GeForce RTX 3080 10GBQ4_K_Moffload5 tok/s$4959.5
NVIDIA GeForce RTX 5060 Ti 8GBQ4_K_Moffload4 tok/s$5007.8
NVIDIA GeForce RTX 4060 Ti 8GBQ4_K_Moffload4 tok/s$5366.9
NVIDIA GeForce RTX 4070Q4_K_Moffload5 tok/s$5709.3
AMD Radeon RX 7600 XTQ4_K_Moffload6 tok/s$5999.5
NVIDIA GeForce RTX 5060 Ti 16GBQ4_K_Moffload8 tok/s$61812.6
AMD Radeon RX 9060 XT 16GBQ4_K_Moffload6 tok/s$6609.1
AMD Radeon RX 7800 XTQ4_K_Moffload8 tok/s$68011.9
NVIDIA GeForce RTX 4060 Ti 16GBQ4_K_Moffload6 tok/s$7308.5
NVIDIA GeForce RTX 4070 SuperQ4_K_Moffload5 tok/s$7407.2
NVIDIA RTX PRO 2000 BlackwellQ4_K_Moffload6 tok/s$7947.8
AMD Radeon RX 9070Q4_K_Moffload8 tok/s$8509.6
NVIDIA GeForce RTX 5070Q4_K_Moffload6 tok/s$8606.5
AMD Radeon RX 9070 XTQ4_K_Moffload8 tok/s$9208.9
NVIDIA GeForce RTX 4070 Ti SuperQ4_K_Moffload9 tok/s$1,1398.0
NVIDIA GeForce RTX 5070 TiQ4_K_Moffload10 tok/s$1,4007.1
NVIDIA GeForce RTX 4080 SuperQ4_K_Moffload9 tok/s$1,4696.3
NVIDIA GeForce RTX 5080Q4_K_Moffload10 tok/s$2,0005.1

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).

QuantFile@8k@32k@64k@128k
Q4_K_Mgguf16.0 GB17.4 GB19.1 GB21.2 GB25.6 GB
Q5_K_Mgguf19.3 GB20.7 GB22.3 GB24.5 GB28.9 GB
Q6_Kgguf22.0 GB23.4 GB25.0 GB27.2 GB31.6 GB
Q8_0gguf26.7 GB28.1 GB29.7 GB31.9 GB36.3 GB

Run it

Copy a command for the runtime you use.

  • Q4_K_M
    • Ollamaollama run hf.co/bartowski/BAAI_AREX-2-GGUF:Q4_K_M
    • LM Studiolms get bartowski/BAAI_AREX-2-GGUF@q4_k_m
    • llama.cppllama-server -hf bartowski/BAAI_AREX-2-GGUF:Q4_K_M
  • Q5_K_M
    • Ollamaollama run hf.co/bartowski/BAAI_AREX-2-GGUF:Q5_K_M
    • LM Studiolms get bartowski/BAAI_AREX-2-GGUF@q5_k_m
    • llama.cppllama-server -hf bartowski/BAAI_AREX-2-GGUF:Q5_K_M
  • Q6_K
    • Ollamaollama run hf.co/bartowski/BAAI_AREX-2-GGUF:Q6_K
    • LM Studiolms get bartowski/BAAI_AREX-2-GGUF@q6_k
    • llama.cppllama-server -hf bartowski/BAAI_AREX-2-GGUF:Q6_K
  • Q8_0
    • Ollamaollama run hf.co/bartowski/BAAI_AREX-2-GGUF:Q8_0
    • LM Studiolms get bartowski/BAAI_AREX-2-GGUF@q8_0
    • llama.cppllama-server -hf bartowski/BAAI_AREX-2-GGUF:Q8_0

Notes

Dense 27B multimodal model fine-tuned from Qwen3.8-27B; requires a recent Transformers release with Qwen3.8 support. The card says to follow the Qwen base model's terms and notices.

Questions

How much VRAM does AREX 2 need?
At Q4_K_M, AREX 2 needs about 17 GB of memory at 8k context and 19 GB at 32k, counting the weights, the KV cache and runtime buffers.
What is the best GPU for running AREX 2 locally?
The cheapest hardware in the US that runs AREX 2 fully in memory at 32k context is the AMD Radeon RX 7900 XT, $926 at Amazon, at about 23 tok/s (estimated). The fastest we price is the NVIDIA GeForce RTX 5090, at about 36 tok/s.
What is the cheapest way to run AREX 2 in the US today?
The AMD Radeon RX 7900 XT, $849 used at eBay, in stock: it runs AREX 2 at Q4_K_M fully in memory at 32k context, at about 23 tok/s (estimated).
Can I run AREX 2 on a 16 GB GPU?
Only partly: at 32k context it does not fit in 16 GB, so part of it runs from system RAM and it slows to about 8 tok/s on the Intel Arc A770 16GB (estimated).
Can I run AREX 2 on a 24 GB GPU?
Yes: at Q5_K_M it fits fully in 24 GB at 32k context. On the NVIDIA GeForce RTX 3090 Ti it runs at about 27 tok/s (estimated).
How do I run AREX 2 locally?
With Ollama: ollama run hf.co/bartowski/BAAI_AREX-2-GGUF:Q4_K_M (the Q4_K_M build). LM Studio and llama.cpp commands are on this page.
What licence is AREX 2 released under?
Apache License 2.0, which allows commercial use.

More hardware for local AI