RX 6900 XT

AMD Radeon RX 6000 (RDNA 2)

VRAM: 16 GB
Memory bandwidth: 512 GB/s
Class: Consumer

Figures assume this GPU plus 32 GB of system RAM (a typical desktop pairing). "What runs on it" is judged at a 8,192-token context. Speeds are estimates, not measurements.

Running local LLMs on AMD Radeon — read this first

Non-NVIDIA cards run today's models, but the software stack is less mature. The practical picture as of July 2026:

ROCm on Windows lags Linux. On Windows a Vulkan backend typically reaches only ~80–90% of the tokens/sec you'd get from ROCm on Linux. For the best AMD performance, run Linux with ROCm.
No FlashAttention-3, NVLink, TensorRT-LLM or ExLlamaV2. Those NVIDIA-only accelerations don't exist here, so expect lower prompt-processing throughput and no multi-GPU NVLink pooling.
The estimated generation speeds on this page come from the memory-bandwidth roofline and hold across vendors; the caveats above mostly affect prompt processing and features, not raw decode tok/s.

Guidance reviewed July 2026 — the ROCm/Vulkan gap narrows over time.

What runs on it

Model	Sweet-spot quant	Est. speed	Community
DeepSeek-R1-Distill-Llama-8B deepseek-ai	Q4_K_M Runs fully on GPU @ 8K ctx	51–68 tok/s (estimate)	no community data
DeepSeek-R1-Distill-Qwen-1.5B deepseek-ai	Q4_K_M Runs fully on GPU @ 8K ctx	227–303 tok/s (estimate)	no community data
DeepSeek-R1-Distill-Qwen-14B deepseek-ai	Q4_K_M Runs fully on GPU @ 8K ctx	29–39 tok/s (estimate)	no community data
DeepSeek-R1-Distill-Qwen-7B deepseek-ai	Q4_K_M Runs fully on GPU @ 8K ctx	60–79 tok/s (estimate)	no community data
Mistral-Small-3.2-24B-Instruct-2506 mistralai	Q4_K_M Runs fully on GPU @ 8K ctx	21–29 tok/s (estimate)	no community data
Qwen2.5-7B-Instruct Qwen	Q4_K_M Runs fully on GPU @ 8K ctx	28–37 tok/s (estimate) pp: 3,791.5 tok/s measured	no community data
Qwen3-8B Qwen	Q4_K_M Runs fully on GPU @ 8K ctx	49–66 tok/s (estimate)	no community data
gemma-4-12B-it google	Q4_K_M Runs fully on GPU @ 8K ctx	43–58 tok/s (estimate)	no community data
gemma-4-E2B-it google	Q4_K_M Runs fully on GPU @ 8K ctx	99–132 tok/s (estimate)	no community data
gemma-4-E4B-it google	Q4_K_M Runs fully on GPU @ 8K ctx	62–82 tok/s (estimate)	no community data
gpt-oss-20b openai	F16 Runs fully on GPU @ 8K ctx	88–118 tok/s (estimate)	no community data
DeepSeek-R1-Distill-Qwen-32B deepseek-ai	Q4_K_M CPU offload @ 8K ctx · ~41 GPU layers	4–6 tok/s (estimate)	no community data
Qwen3-30B-A3B-Instruct-2507 Qwen	Q4_K_M CPU offload @ 8K ctx · ~36 GPU layers	42–56 tok/s (estimate)	no community data
Qwen3-32B Qwen	Q4_K_M CPU offload @ 8K ctx · ~41 GPU layers	4–6 tok/s (estimate)	no community data
Qwen3-Coder-30B-A3B-Instruct Qwen	Q4_K_M CPU offload @ 8K ctx · ~36 GPU layers	42–56 tok/s (estimate)	no community data
Qwen3.6-27B Qwen	Q4_K_M CPU offload @ 8K ctx · ~49 GPU layers	7–9 tok/s (estimate)	no community data
Qwen3.6-35B-A3B Qwen	Q4_K_M CPU offload @ 8K ctx · ~25 GPU layers	4–5 tok/s (estimate)	no community data
gemma-4-26B-A4B-it google	Q4_K_M CPU offload @ 8K ctx · ~23 GPU layers	7–9 tok/s (estimate)	no community data
gemma-4-31B-it google	Q4_K_M CPU offload @ 8K ctx · ~0 GPU layers	2–3 tok/s (estimate)	no community data

"Est. speed" is a modelled range labelled estimate (D8) for generation (decode) throughput. "Community" shows the median of approved user-submitted reports on this GPU class only where enough exist — never an estimate. "pp" is measured prompt-processing (ingestion) throughput from approved community reports; rows without a measurement show none.