RX 7900 XT 20GB

AMD · 20 GB GDDR6 · 800 GB/s · 315 W · MSRP $899 · Search Amazon (affiliate)

The RX 7900 XT 20GB is a AMD accelerator with 20 GB GDDR6 and about 800 GB/s peak memory bandwidth. AI Hashrate estimates local LLM decode (batch 1) at 8K context: fit counts weights + KV cache + 1 GB runtime overhead — not weights alone. At Q4, about 32 curated models fit fully on this card; at FP16, about 15 fit. Top Q4 speeds: Qwen3-0.6B ≈ 636.4 tok/s (estimated, fits); gemma-3-1b-it ≈ 509.1 tok/s (estimated, fits); Llama-3.2-1B-Instruct ≈ 424.2 tok/s (estimated, fits); Qwen3-1.7B ≈ 254.5 tok/s (estimated, fits); Qwen3-30B-A3B ≈ 169.7 tok/s (estimated, fits). Relative ranking is more reliable than absolute tok/s. See methodology for the bandwidth formula and measured-anchor policy. Methodology.

Table: decode speed estimates (Q4 / FP16). Measured rows override estimates. MSRP is list price, not live retail. Context default 8K.

Model	Params (B)	Quant	tok/s	Fits?
Qwen3-0.6B	0.8	Q4	636.4 est.	Yes
gemma-3-1b-it	1.0	Q4	509.1 est.	Yes
Llama-3.2-1B-Instruct	1.2	Q4	424.2 est.	Yes
Qwen3-1.7B	2.0	Q4	254.5 est.	Yes
Qwen3-0.6B	0.8	FP16	175.0 est.	Yes
Qwen3-30B-A3B	30.5	Q4	169.7 est.	Yes
Qwen3-Coder-30B-A3B-Instruct	30.5	Q4	169.7 est.	Yes
GLM-4.7-Flash	31.2	Q4	169.7 est.	Yes
Qwen3-30B-A3B-Instruct-2507	30.5	Q4	169.7 est.	Yes
Llama-3.2-3B-Instruct	3.2	Q4	159.1 est.	Yes
gpt-oss-20b	21.5	Q4	141.4 est.	Yes
gemma-3-1b-it	1.0	FP16	140.0 est.	Yes
Phi-4-mini-instruct	3.8	Q4	134.0 est.	Yes
Phi-4-mini-reasoning	3.8	Q4	134.0 est.	Yes
gemma-3-4b-it	4.3	Q4	130.5 est.	Yes
Qwen3-4B-Instruct-2507	4.0	Q4	127.3 est.	Yes
Qwen3-4B-Thinking-2507	4.0	Q4	127.3 est.	Yes
Llama-3.2-1B-Instruct	1.2	FP16	116.7 est.	Yes
Qwen3-Coder-Next	79.7	Q4	93.9 est.	No
Qwen3-Next-80B-A3B-Instruct	81.3	Q4	90.4 est.	No
Mistral-7B-Instruct-v0.3	7.2	Q4	70.7 est.	Yes
Qwen3-1.7B	2.0	FP16	70.0 est.	Yes
Qwen2.5-7B-Instruct	7.6	Q4	67.0 est.	Yes
Qwen2.5-Coder-7B-Instruct	7.6	Q4	67.0 est.	Yes
DeepSeek-R1-Distill-Qwen-7B	7.6	Q4	67.0 est.	Yes
Llama-3.1-8B-Instruct	8.0	Q4	63.6 est.	Yes
Qwen3-8B	8.2	Q4	62.1 est.	Yes
DeepSeek-R1-0528-Qwen3-8B	8.2	Q4	62.1 est.	Yes
Qwen3-8B	8.2	FP16	53.8 est.	No
DeepSeek-R1-0528-Qwen3-8B	8.2	FP16	53.8 est.	No
Mixtral-8x7B-Instruct-v0.1	46.7	Q4	52.6 est.	No
gemma-3-27b-it	27.4	Q4	46.2 est.	No
Llama-3.2-3B-Instruct	3.2	FP16	43.8 est.	Yes
gemma-3-12b-it	12.2	Q4	42.4 est.	Yes
Phi-4-mini-instruct	3.8	FP16	36.8 est.	Yes
Phi-4-mini-reasoning	3.8	FP16	36.8 est.	Yes
gemma-3-4b-it	4.3	FP16	35.9 est.	Yes
Qwen3-4B-Instruct-2507	4.0	FP16	35.0 est.	Yes
Qwen3-4B-Thinking-2507	4.0	FP16	35.0 est.	Yes
Phi-4	14.7	Q4	34.6 est.	Yes
Phi-4-reasoning	14.7	Q4	34.6 est.	Yes
Qwen3-14B	14.8	Q4	34.4 est.	Yes
Qwen2.5-14B-Instruct	14.8	Q4	34.4 est.	Yes
Qwen2.5-Coder-14B-Instruct	14.8	Q4	34.4 est.	Yes
DeepSeek-R1-Distill-Qwen-14B	14.8	Q4	34.4 est.	Yes
Qwen3-32B	32.8	Q4	27.1 est.	No
Qwen2.5-32B-Instruct	32.8	Q4	27.1 est.	No
Qwen2.5-Coder-32B-Instruct	32.8	Q4	27.1 est.	No
DeepSeek-R1-Distill-Qwen-32B	32.8	Q4	27.1 est.	No
gpt-oss-120b	120.4	Q4	24.5 est.	No
gpt-oss-20b	21.5	FP16	22.2 est.	No
Devstral-Small-2507	23.6	Q4	21.6 est.	Yes
Mistral-Small-3.2-24B-Instruct-2506	24.0	Q4	21.2 est.	Yes
Magistral-Small-2509	24.0	Q4	21.2 est.	Yes
Mistral-7B-Instruct-v0.3	7.2	FP16	19.4 est.	Yes
Qwen2.5-7B-Instruct	7.6	FP16	18.4 est.	Yes
Qwen2.5-Coder-7B-Instruct	7.6	FP16	18.4 est.	Yes
DeepSeek-R1-Distill-Qwen-7B	7.6	FP16	18.4 est.	Yes
Llama-3.1-8B-Instruct	8.0	FP16	17.5 est.	Yes
gemma-3-12b-it	12.2	FP16	17.3 est.	No
Qwen3-30B-A3B	30.5	FP16	13.6 est.	No
Qwen3-Coder-30B-A3B-Instruct	30.5	FP16	13.6 est.	No
Qwen3-30B-A3B-Instruct-2507	30.5	FP16	13.6 est.	No
GLM-4.7-Flash	31.2	FP16	13.0 est.	No
GLM-4.5-Air	110.5	Q4	11.8 est.	No
Phi-4	14.7	FP16	9.8 est.	No
Phi-4-reasoning	14.7	FP16	9.8 est.	No
Qwen3-14B	14.8	FP16	9.6 est.	No
Qwen2.5-14B-Instruct	14.8	FP16	9.6 est.	No
Qwen2.5-Coder-14B-Instruct	14.8	FP16	9.6 est.	No
DeepSeek-R1-Distill-Qwen-14B	14.8	FP16	9.6 est.	No
Llama-3.1-70B-Instruct	70.6	Q4	2.9 est.	No
Llama-3.3-70B-Instruct	70.6	Q4	2.9 est.	No
DeepSeek-R1-Distill-Llama-70B	70.6	Q4	2.8 est.	No
Qwen2.5-72B-Instruct	72.7	Q4	2.6 est.	No
Devstral-Small-2507	23.6	FP16	2.4 est.	No
Mistral-Small-3.2-24B-Instruct-2506	24.0	FP16	2.3 est.	No
Magistral-Small-2509	24.0	FP16	2.3 est.	No
gemma-3-27b-it	27.4	FP16	1.6 est.	No
Qwen3-32B	32.8	FP16	0.9 est.	No
Qwen2.5-32B-Instruct	32.8	FP16	0.9 est.	No
Qwen2.5-Coder-32B-Instruct	32.8	FP16	0.9 est.	No
DeepSeek-R1-Distill-Qwen-32B	32.8	FP16	0.9 est.	No

Fit checks

• Qwen3-0.6B — fits · ≈636.4 tok/s Q4 est.
• Qwen3-8B — fits · ≈62.1 tok/s Q4 est.
• Qwen2.5-7B-Instruct — fits · ≈67.0 tok/s Q4 est.
• Llama-3.2-1B-Instruct — fits · ≈424.2 tok/s Q4 est.
• Llama-3.1-8B-Instruct — fits · ≈63.6 tok/s Q4 est.
• gpt-oss-20b — fits · ≈141.4 tok/s Q4 est.
• Qwen3-32B — does not fit · ≈27.1 tok/s Q4 est.
• Qwen3-1.7B — fits · ≈254.5 tok/s Q4 est.
• Mistral-7B-Instruct-v0.3 — fits · ≈70.7 tok/s Q4 est.
• Qwen3-14B — fits · ≈34.4 tok/s Q4 est.
• Qwen3-4B-Instruct-2507 — fits · ≈127.3 tok/s Q4 est.

All 61 model fit checks — see the linked Fits? column in the table above