FitMyLLM
▸ AMD· GCN 4.0

AMD Radeon Pro Duo Polaris

Running LLMs on the Radeon Pro Duo Polaris — the long read: which models fit at which quantisation, and the settings worth changing. · Or what a budget buys

VRAM
16 GB
MID-RANGE
BANDWIDTH
224
GB/S
MODELS Q4
262/449
58%
7B Q4 SPEED
~28
GOOD
▸ MODEL COVERAGE @ Q458% OF ALL
▸ ESTIMATED SPEED· BY MODEL SIZE @ Q4

Average speeds at Q4 quantization. Actual performance varies by model architecture and context length.

3B
~66
TOK/S
7B
~28
TOK/S
14B
~14
TOK/S
32B
—
18.0GB NEEDED
70B
—
39.4GB NEEDED
▸ MEASURED RIG REPORTS

We rent the machine and time every model on it: decode, VRAM peak, concurrency, watts and cost per million tokens. Yours may already be one of them — and two are free to read in full.

SEE THE REPORTS →
▸ SPECIFICATIONS
VRAM
16 GB
BANDWIDTH
224 GB/s
FP16 COMPUTE
5.7 TFLOPS
TDP
250W
MEMORY
GDDR5
ARCHITECTURE
GCN 4.0
STREAM PROCESSORS
2,304
COMPUTE UNITS
36
PCIE
Gen 3 x16
124
FAST MODELS · >30 TOK/S
Real-time chat speed
262
USABLE · >10 TOK/S
Comfortable for all tasks
262
TOTAL COMPATIBLE
Fit in VRAM at Q4
▸ DON’T WANT TO BUY?

Test Radeon Pro Duo Polaris (or anything bigger) without committing. Pay by the second, cancel anytime.

Rent on Vast.ai →Or RunPod →

Spin up in ~60s. Pay by the second. Cancel anytime.

Some links are affiliate links — we may earn a small commission at no extra cost to you. This helps keep FitMyLLM free and independent.

▸ COMPATIBLE MODELS· 262
S
Falcon-H1R Tiny 90M0.09B
FALCON·256K CTX· CHAT· REASONING
2212
TOK/S · 3% VRAM
›
S
SmolLM2 135M0.135B
SMOLLM·2K CTX· CHAT
1475
TOK/S · 4% VRAM
›
S
nomic-embed-text-v1.5 100M0.14B
EMBEDDING·8K CTX· CHAT
1422
TOK/S · 4% VRAM
›
S
GPT-2 124M0.14B
GPT2·1K CTX· CHAT
1422
TOK/S · 4% VRAM
›
S
SmolVLM 256M0.256B
SMOLLM·8K CTX· CHAT· VISION
778
TOK/S · 4% VRAM
›
S
Gemma 3 270M0.27B
GEMMA·32K CTX· CHAT
737
TOK/S · 4% VRAM
›
S
Snowflake Arctic Embed M v2.00.305B
EMBEDDING·8K CTX· EMBEDDING· MULTILINGUAL
653
TOK/S · 4% VRAM
›
S
bge-large-en-v1.5 335M0.335B
EMBEDDING·1K CTX· CHAT
594
TOK/S · 4% VRAM
›
S
mxbai-embed-large-v10.335B
EMBEDDING·1K CTX· EMBEDDING
594
TOK/S · 4% VRAM
›
S
Snowflake Arctic Embed L0.335B
EMBEDDING·1K CTX· EMBEDDING
594
TOK/S · 4% VRAM
›
S
LFM2 350M0.35B
LFM·125K CTX· CHAT· TOOL_USE
569
TOK/S · 4% VRAM
›
S
SmolLM2 360M0.36B
SMOLLM·8K CTX· CHAT
553
TOK/S · 4% VRAM
›
S
GPT-2 Medium 345M0.38B
GPT2·1K CTX· CHAT
524
TOK/S · 5% VRAM
›
S
Qwen 2.5 0.5B0.5B
QWEN·32K CTX· CHAT
398
TOK/S · 5% VRAM
›
S
SmolVLM 500M0.5B
SMOLLM·8K CTX· CHAT· VISION
398
TOK/S · 5% VRAM
›
S
Falcon-H1 0.5B0.52B
FALCON·128K CTX· CHAT
383
TOK/S · 5% VRAM
›
S
BGE-M30.568B
EMBEDDING·8K CTX· EMBEDDING
351
TOK/S · 5% VRAM
›
S
Snowflake Arctic Embed L v2.00.568B
EMBEDDING·8K CTX· EMBEDDING· MULTILINGUAL
351
TOK/S · 5% VRAM
›
S
Qwen3-Embedding 0.6B0.6B
EMBEDDING·32K CTX· EMBEDDING· MULTILINGUAL
332
TOK/S · 5% VRAM
›
S
Falcon-H1R Tiny 0.6B0.6B
FALCON·32K CTX· CHAT· REASONING
332
TOK/S · 5% VRAM
›
S
Falcon Perception 0.6B0.6B
FALCON·4K CTX· VISION
332
TOK/S · 5% VRAM
›
S
Qwen 1.5 0.5B0.62B
QWEN·32K CTX· CHAT
321
TOK/S · 5% VRAM
›
S
LFM2 700M0.74B
LFM·125K CTX· CHAT· TOOL_USE
269
TOK/S · 6% VRAM
›
S
Qwen3 0.6B0.75B
QWEN·32K CTX· CHAT· REASONING
265
TOK/S · 6% VRAM
›
S
GPT-2 Large 774M0.81B
GPT2·1K CTX· CHAT
246
TOK/S · 6% VRAM
›
S
Qwen 3.5 0.8B0.87B
QWEN·256K CTX· CHAT· CODING· MULTILINGUAL
229
TOK/S · 6% VRAM
›
S
InternVL3 1B0.94B
OTHER·32K CTX· CHAT· VISION
212
TOK/S · 7% VRAM
›
S
MiniCPM5 1B1.08B
MINICPM·128K CTX· CHAT· REASONING· MULTILINGUAL
184
TOK/S · 7% VRAM
›
S
TinyLlama 1.1B1.1B
LLAMA·2K CTX· CHAT
181
TOK/S · 7% VRAM
›
S
LFM2.5-1.2B-Thinking1.2B
LFM·122K CTX· CHAT· REASONING· TOOL_USE
166
TOK/S · 8% VRAM
›
S
Llama-3.2-1B1.2B
LLAMA·4K CTX· CHAT
166
TOK/S · 8% VRAM
›
S
LFM2 1.2B1.2B
LFM·125K CTX· CHAT· TOOL_USE· MULTILINGUAL
166
TOK/S · 8% VRAM
›
S
Zamba2 1.2B1.2B
OTHER·4K CTX· CHAT
166
TOK/S · 8% VRAM
›
S
EXAONE-4.0-1.2B1.3B
EXAONE·64K CTX· CHAT
153
TOK/S · 8% VRAM
›
S
OPT 1.3B1.3B
OPT·2K CTX· CHAT
153
TOK/S · 8% VRAM
›
S
MiniCPM-V 4.61.3B
OTHER·256K CTX· CHAT· VISION
153
TOK/S · 8% VRAM
›
S
DeepSeek Coder 1.3B1.35B
DEEPSEEK·16K CTX· CODING
147
TOK/S · 8% VRAM
›
S
Phi-1 1.3B1.42B
PHI·2K CTX· CODING
140
TOK/S · 8% VRAM
›
S
Phi-1.5 1.3B1.42B
PHI·2K CTX· CHAT· CODING
140
TOK/S · 8% VRAM
›
S
LFM2 8B A1B8.3BMoE
LFM·125K CTX· CHAT· TOOL_USE· MULTILINGUAL
133
TOK/S · 35% VRAM
›
▸ NEXT STEP

Get personalized recommendations.

See ranked models with benchmark scores, run commands, and precise speed estimates for your Radeon Pro Duo Polaris.

▸Find models for Radeon Pro Duo PolarisCompare with other GPUs
WHAT THIS CARD IS WORTH

Radeon Pro Duo Polaris holds 262 of the models in our catalogue and is, in practice, a IQ4_XS card — the largest it takes is ERNIE 4.5 21B A3B at IQ4_XS.

TOKENS/SEC PER $100
—
8B at Q4_K_M, so cards compare like for like
VRAM PER $100
—
what memory costs on this card
THE BIGGEST IT TAKES
ERNIE 4.5 21B A3B21.95B · IQ4_XS13.9 GB52 tok/sEST
GPT-OSS 20B21B · Q4_K_S13.8 GB44 tok/sEST
Reka Flash 321B · Q4_K_S14.2 GB11 tok/sEST
Reka Flash 3.121B · Q4_K_S14.2 GB11 tok/sEST
InternLM2 20B19.8B · Q4_K_M14.3 GB11 tok/sEST
InternLM2.5 20B19.8B · Q4_K_M14.3 GB11 tok/sEST
Ling-lite 16.8B16.8B · Q5_K_M13.6 GB52 tok/sEST
DeepSeek V2 Lite 16B16B · Q5_K_M12.9 GB52 tok/sEST
AGAINST CARDS OF SIMILAR MEMORY
M1 Max (32GB)21 GB$1,4993.0 tok/s per $100
M2 Max (32GB)21 GB$1,7992.5 tok/s per $100
M2 Pro (32GB)21 GB$1,4991.9 tok/s per $100
M4 (32GB)21 GB$1,1991.7 tok/s per $100
Radeon Pro Duo Polaris16 GB——

Neighbours in memory rather than in price: memory decides whether a card can do the job at all, so two cards of the same size at different prices is the comparison you are making. Ordered by memory, then bandwidth — not by the value column, which is worked out from bandwidth and price alone and therefore rewards a cheap card whatever its software stack does to that bandwidth in practice. Read it as one input, not as a ranking.

▸ DEVICE UNDER TEST

AMD Radeon Pro Duo Polaris — 16 GB VRAM.

▸ RADEON PRO DUO POLARIS SPEC
BRAND
AMD
VRAM
16 GB GDDR5
BANDWIDTH
224 GB/s
FP16 COMPUTE
5.7 TFLOPS
FP32 COMPUTE
5.7 TFLOPS
STREAM PROCESSORS
2,304
TDP
250 W
ARCHITECTURE
GCN 4.0
▸ AI CAPABILITY
262/ 449 models @ Q4

With 16 GB VRAM and 224 GB/s bandwidth, this GPU handles models up to 19.8B parameters.

Speed ≈ bandwidth / model_size × efficiency. A 7B model at Q4 runs at ~26 tok/s.

RENT IT — LIVE PRICES
Checking what the clouds are charging…
§ 01TOP MODELS FOR RADEON PRO DUO POLARIS
262 FIT · SHOWING 20
MODELSIZEVRAM Q4TOK/SAVG
InternLM2 20B19.8B12.6 GB1045.1
InternLM2.5 20B19.8B12.6 GB1050.9
Ling-lite 16.8B16.8B10.8 GB83—
DeepSeek V2 Lite 16B16B10.3 GB8338.0
StarCoder2 15B15.96B10.2 GB1226.5
DeepSeek-Coder-V2-Lite 15.7B15.7B10.1 GB8343.0
DeepSeek-VL2 Small 16B15.7B10.1 GB8343.1
StarCoder 15B15.5B10.0 GB1321.0
InternVL3 14B15.12B9.7 GB1338.1
Phi-4-reasoning-vision 15B15B9.7 GB1342.8
DeepSeek R1 Distill Qwen 14B14.8B9.5 GB1343.9
DeepCoder 14B14.8B9.5 GB1338.7
Qwen2.5-Coder-14B14.8B9.5 GB1341.3
Qwen2.5-14B14.8B9.5 GB1341.3
Qwen3 14B14.8B9.5 GB1345.7
phi-4 14B14.66B9.4 GB1433.7
Phi-4-reasoning 14B14.66B9.4 GB1433.7
Phi-4-reasoning-plus 14B14.66B9.4 GB1475.5
Ministral 3 14B14B9.0 GB1425.9
Phi-3-medium-14b14B9.0 GB1433.7