RTX 4070 12GB — local LLM fit & speed

NVIDIA · 12 GB GDDR6X · 504 GB/s · 200 W · MSRP $599 · ASUS Dual GeForce RTX 4070 OC 12GB (affiliate) · More listings

The RTX 4070 12GB is a NVIDIA accelerator with 12 GB GDDR6X and about 504 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 ≈ 400.9 tok/s (estimated, fits); gemma-3-1b-it ≈ 320.7 tok/s (estimated, fits); Llama-3.2-1B-Instruct ≈ 267.3 tok/s (estimated, fits); Qwen3-1.7B ≈ 160.4 tok/s (estimated, fits); Llama-3.2-3B-Instruct ≈ 100.2 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	400.9 est.	Yes
gemma-3-1b-it	1.0	Q4	320.7 est.	Yes
Llama-3.2-1B-Instruct	1.2	Q4	267.3 est.	Yes
gpt-oss-20b	21.5	Q4	199.6 est.	No
Qwen3-1.7B	2.0	Q4	160.4 est.	Yes
Qwen3-30B-A3B	30.5	Q4	130.2 est.	No
Qwen3-Coder-30B-A3B-Instruct	30.5	Q4	130.2 est.	No
Qwen3-30B-A3B-Instruct-2507	30.5	Q4	130.2 est.	No
Qwen3-0.6B	0.8	FP16	110.2 est.	Yes
Llama-3.2-3B-Instruct	3.2	Q4	100.2 est.	Yes
gemma-3-1b-it	1.0	FP16	88.2 est.	Yes
Phi-4-mini-instruct	3.8	Q4	84.4 est.	Yes
Phi-4-mini-reasoning	3.8	Q4	84.4 est.	Yes
gemma-3-4b-it	4.3	Q4	82.2 est.	Yes
Qwen3-4B-Instruct-2507	4.0	Q4	80.2 est.	Yes
Qwen3-4B-Thinking-2507	4.0	Q4	80.2 est.	Yes
Llama-3.2-1B-Instruct	1.2	FP16	73.5 est.	Yes
Phi-4	14.7	Q4	62.0 est.	No
Phi-4-reasoning	14.7	Q4	62.0 est.	No
Qwen3-14B	14.8	Q4	60.9 est.	No
Qwen2.5-14B-Instruct	14.8	Q4	60.9 est.	No
Qwen2.5-Coder-14B-Instruct	14.8	Q4	60.9 est.	No
DeepSeek-R1-Distill-Qwen-14B	14.8	Q4	60.9 est.	No
Llama-3.1-8B-Instruct	8.0	Q4	52.1 measured (singhajit-hwcorner)	Yes
Mistral-7B-Instruct-v0.3	7.2	Q4	44.5 est.	Yes
Qwen3-1.7B	2.0	FP16	44.1 est.	Yes
Qwen2.5-7B-Instruct	7.6	Q4	42.2 est.	Yes
Qwen2.5-Coder-7B-Instruct	7.6	Q4	42.2 est.	Yes
DeepSeek-R1-Distill-Qwen-7B	7.6	Q4	42.2 est.	Yes
Qwen3-8B	8.2	Q4	39.1 est.	Yes
DeepSeek-R1-0528-Qwen3-8B	8.2	Q4	39.1 est.	Yes
Llama-3.2-3B-Instruct	3.2	FP16	27.6 est.	Yes
gemma-3-12b-it	12.2	Q4	26.7 est.	Yes
GLM-4.7-Flash	31.2	Q4	26.0 est.	No
Phi-4-mini-instruct	3.8	FP16	23.2 est.	Yes
Phi-4-mini-reasoning	3.8	FP16	23.2 est.	Yes
gemma-3-4b-it	4.3	FP16	22.6 est.	Yes
Qwen3-4B-Instruct-2507	4.0	FP16	22.0 est.	Yes
Qwen3-4B-Thinking-2507	4.0	FP16	22.0 est.	Yes
Qwen3-Coder-Next	79.7	Q4	21.3 est.	No
Qwen3-Next-80B-A3B-Instruct	81.3	Q4	20.5 est.	No
Mistral-7B-Instruct-v0.3	7.2	FP16	17.8 est.	No
Devstral-Small-2507	23.6	Q4	16.0 est.	No
Qwen2.5-7B-Instruct	7.6	FP16	15.2 est.	No
Qwen2.5-Coder-7B-Instruct	7.6	FP16	15.2 est.	No
DeepSeek-R1-Distill-Qwen-7B	7.6	FP16	15.2 est.	No
Mistral-Small-3.2-24B-Instruct-2506	24.0	Q4	15.2 est.	No
Magistral-Small-2509	24.0	Q4	15.2 est.	No
Llama-3.1-8B-Instruct	8.0	FP16	13.1 est.	No
Qwen3-8B	8.2	FP16	12.2 est.	No
DeepSeek-R1-0528-Qwen3-8B	8.2	FP16	12.2 est.	No
Mixtral-8x7B-Instruct-v0.1	46.7	Q4	11.9 est.	No
gemma-3-27b-it	27.4	Q4	10.5 est.	No
Qwen3-32B	32.8	Q4	6.1 est.	No
Qwen2.5-32B-Instruct	32.8	Q4	6.1 est.	No
Qwen2.5-Coder-32B-Instruct	32.8	Q4	6.1 est.	No
DeepSeek-R1-Distill-Qwen-32B	32.8	Q4	6.1 est.	No
gpt-oss-20b	21.5	FP16	5.0 est.	No
gemma-3-12b-it	12.2	FP16	3.9 est.	No
Qwen3-14B	14.8	FP16	2.2 est.	No
Qwen2.5-14B-Instruct	14.8	FP16	2.2 est.	No
Qwen2.5-Coder-14B-Instruct	14.8	FP16	2.2 est.	No
DeepSeek-R1-Distill-Qwen-14B	14.8	FP16	2.2 est.	No
Phi-4	14.7	FP16	2.2 est.	No
Phi-4-reasoning	14.7	FP16	2.2 est.	No