RTX 5070 12GB — local LLM fit & speed

NVIDIA · 12 GB GDDR7 · 672 GB/s · 250 W · MSRP $549 · Search Amazon (affiliate)

The RTX 5070 12GB is a NVIDIA accelerator with 12 GB GDDR7 and about 672 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 18 curated models fit fully on this card; at FP16, about 10 fit. Top Q4 speeds: Qwen3-0.6B ≈ 534.5 tok/s (estimated, fits); gemma-3-1b-it ≈ 427.6 tok/s (estimated, fits); Llama-3.2-1B-Instruct ≈ 356.4 tok/s (estimated, fits); Qwen3-1.7B ≈ 213.8 tok/s (estimated, fits); Llama-3.2-3B-Instruct ≈ 133.6 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	534.5 est.	Yes
gemma-3-1b-it	1.0	Q4	427.6 est.	Yes
Llama-3.2-1B-Instruct	1.2	Q4	356.4 est.	Yes
gpt-oss-20b	21.5	Q4	266.2 est.	No
Qwen3-1.7B	2.0	Q4	213.8 est.	Yes
Qwen3-30B-A3B	30.5	Q4	173.6 est.	No
Qwen3-Coder-30B-A3B-Instruct	30.5	Q4	173.6 est.	No
Qwen3-30B-A3B-Instruct-2507	30.5	Q4	173.6 est.	No
Qwen3-0.6B	0.8	FP16	147.0 est.	Yes
Llama-3.2-3B-Instruct	3.2	Q4	133.6 est.	Yes
gemma-3-1b-it	1.0	FP16	117.6 est.	Yes
Phi-4-mini-instruct	3.8	Q4	112.5 est.	Yes
Phi-4-mini-reasoning	3.8	Q4	112.5 est.	Yes
gemma-3-4b-it	4.3	Q4	109.7 est.	Yes
Qwen3-4B-Instruct-2507	4.0	Q4	106.9 est.	Yes
Qwen3-4B-Thinking-2507	4.0	Q4	106.9 est.	Yes
Llama-3.2-1B-Instruct	1.2	FP16	98.0 est.	Yes
Phi-4	14.7	Q4	82.7 est.	No
Phi-4-reasoning	14.7	Q4	82.7 est.	No
Qwen3-14B	14.8	Q4	81.2 est.	No
Qwen2.5-14B-Instruct	14.8	Q4	81.2 est.	No
Qwen2.5-Coder-14B-Instruct	14.8	Q4	81.2 est.	No
DeepSeek-R1-Distill-Qwen-14B	14.8	Q4	81.2 est.	No
Mistral-7B-Instruct-v0.3	7.2	Q4	59.4 est.	Yes
Qwen3-1.7B	2.0	FP16	58.8 est.	Yes
Qwen2.5-7B-Instruct	7.6	Q4	56.3 est.	Yes
Qwen2.5-Coder-7B-Instruct	7.6	Q4	56.3 est.	Yes
DeepSeek-R1-Distill-Qwen-7B	7.6	Q4	56.3 est.	Yes
Llama-3.1-8B-Instruct	8.0	Q4	53.5 est.	Yes
Qwen3-8B	8.2	Q4	52.2 est.	Yes
DeepSeek-R1-0528-Qwen3-8B	8.2	Q4	52.2 est.	Yes
Llama-3.2-3B-Instruct	3.2	FP16	36.8 est.	Yes
gemma-3-12b-it	12.2	Q4	35.6 est.	Yes
GLM-4.7-Flash	31.2	Q4	34.7 est.	No
Phi-4-mini-instruct	3.8	FP16	30.9 est.	Yes
Phi-4-mini-reasoning	3.8	FP16	30.9 est.	Yes
gemma-3-4b-it	4.3	FP16	30.2 est.	Yes
Qwen3-4B-Instruct-2507	4.0	FP16	29.4 est.	Yes
Qwen3-4B-Thinking-2507	4.0	FP16	29.4 est.	Yes
Qwen3-Coder-Next	79.7	Q4	28.4 est.	No
Qwen3-Next-80B-A3B-Instruct	81.3	Q4	27.3 est.	No
Mistral-7B-Instruct-v0.3	7.2	FP16	23.7 est.	No
Devstral-Small-2507	23.6	Q4	21.3 est.	No
Qwen2.5-7B-Instruct	7.6	FP16	20.3 est.	No
Qwen2.5-Coder-7B-Instruct	7.6	FP16	20.3 est.	No
DeepSeek-R1-Distill-Qwen-7B	7.6	FP16	20.3 est.	No
Mistral-Small-3.2-24B-Instruct-2506	24.0	Q4	20.3 est.	No
Magistral-Small-2509	24.0	Q4	20.3 est.	No
Llama-3.1-8B-Instruct	8.0	FP16	17.5 est.	No
Qwen3-8B	8.2	FP16	16.3 est.	No
DeepSeek-R1-0528-Qwen3-8B	8.2	FP16	16.3 est.	No
Mixtral-8x7B-Instruct-v0.1	46.7	Q4	15.9 est.	No
gemma-3-27b-it	27.4	Q4	14.0 est.	No
Qwen3-32B	32.8	Q4	8.2 est.	No
Qwen2.5-32B-Instruct	32.8	Q4	8.2 est.	No
Qwen2.5-Coder-32B-Instruct	32.8	Q4	8.2 est.	No
DeepSeek-R1-Distill-Qwen-32B	32.8	Q4	8.2 est.	No
gpt-oss-20b	21.5	FP16	6.7 est.	No
gemma-3-12b-it	12.2	FP16	5.2 est.	No
Phi-4	14.7	FP16	3.0 est.	No
Phi-4-reasoning	14.7	FP16	3.0 est.	No
Qwen3-14B	14.8	FP16	2.9 est.	No
Qwen2.5-14B-Instruct	14.8	FP16	2.9 est.	No
Qwen2.5-Coder-14B-Instruct	14.8	FP16	2.9 est.	No
DeepSeek-R1-Distill-Qwen-14B	14.8	FP16	2.9 est.	No