L40S 48GB — local LLM fit & speed

NVIDIA · 48 GB GDDR6 · 864 GB/s · 350 W · MSRP $8,000 · Search Amazon (affiliate)

The L40S 48GB is a NVIDIA accelerator with 48 GB GDDR6 and about 864 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 39 curated models fit fully on this card; at FP16, about 25 fit. Top Q4 speeds: Qwen3-0.6B ≈ 687.3 tok/s (estimated, fits); gemma-3-1b-it ≈ 549.8 tok/s (estimated, fits); Llama-3.2-1B-Instruct ≈ 458.2 tok/s (estimated, fits); Qwen3-1.7B ≈ 274.9 tok/s (estimated, fits); Qwen3-30B-A3B ≈ 183.3 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	687.3 est.	Yes
Qwen3-Next-80B-A3B-Instruct	81.3	Q4	562.3 est.	No
gemma-3-1b-it	1.0	Q4	549.8 est.	Yes
Llama-3.2-1B-Instruct	1.2	Q4	458.2 est.	Yes
Qwen3-1.7B	2.0	Q4	274.9 est.	Yes
Qwen3-0.6B	0.8	FP16	189.0 est.	Yes
Qwen3-30B-A3B	30.5	Q4	183.3 est.	Yes
Qwen3-Coder-30B-A3B-Instruct	30.5	Q4	183.3 est.	Yes
Qwen3-30B-A3B-Instruct-2507	30.5	Q4	183.3 est.	Yes
Qwen3-Coder-Next	79.7	Q4	183.3 est.	Yes
Llama-3.2-3B-Instruct	3.2	Q4	171.8 est.	Yes
gpt-oss-20b	21.5	Q4	152.7 est.	Yes
gemma-3-1b-it	1.0	FP16	151.2 est.	Yes
Phi-4-mini-instruct	3.8	Q4	144.7 est.	Yes
Phi-4-mini-reasoning	3.8	Q4	144.7 est.	Yes
gemma-3-4b-it	4.3	Q4	141.0 est.	Yes
Qwen3-4B-Instruct-2507	4.0	Q4	137.5 est.	Yes
Qwen3-4B-Thinking-2507	4.0	Q4	137.5 est.	Yes
Llama-3.2-1B-Instruct	1.2	FP16	126.0 est.	Yes
gpt-oss-120b	120.4	Q4	95.5 est.	No
Qwen3-30B-A3B	30.5	FP16	84.7 est.	No
Qwen3-Coder-30B-A3B-Instruct	30.5	FP16	84.7 est.	No
Qwen3-30B-A3B-Instruct-2507	30.5	FP16	84.7 est.	No
Mistral-7B-Instruct-v0.3	7.2	Q4	76.4 est.	Yes
Qwen3-1.7B	2.0	FP16	75.6 est.	Yes
GLM-4.5-Air	110.5	Q4	73.2 est.	No
Qwen2.5-7B-Instruct	7.6	Q4	72.3 est.	Yes
Qwen2.5-Coder-7B-Instruct	7.6	Q4	72.3 est.	Yes
DeepSeek-R1-Distill-Qwen-7B	7.6	Q4	72.3 est.	Yes
Llama-3.1-8B-Instruct	8.0	Q4	68.7 est.	Yes
Qwen3-8B	8.2	Q4	67.1 est.	Yes
DeepSeek-R1-0528-Qwen3-8B	8.2	Q4	67.1 est.	Yes
Llama-3.2-3B-Instruct	3.2	FP16	47.2 est.	Yes
gemma-3-12b-it	12.2	Q4	45.8 est.	Yes
GLM-4.7-Flash	31.2	Q4	45.8 est.	Yes
Mixtral-8x7B-Instruct-v0.1	46.7	Q4	42.6 est.	Yes
gpt-oss-20b	21.5	FP16	42.0 est.	Yes
Phi-4-mini-instruct	3.8	FP16	39.8 est.	Yes
Phi-4-mini-reasoning	3.8	FP16	39.8 est.	Yes
gemma-3-4b-it	4.3	FP16	38.8 est.	Yes
Qwen3-4B-Instruct-2507	4.0	FP16	37.8 est.	Yes
Qwen3-4B-Thinking-2507	4.0	FP16	37.8 est.	Yes
Phi-4	14.7	Q4	37.4 est.	Yes
Phi-4-reasoning	14.7	Q4	37.4 est.	Yes
Qwen3-14B	14.8	Q4	37.1 est.	Yes
Qwen2.5-14B-Instruct	14.8	Q4	37.1 est.	Yes
Qwen2.5-Coder-14B-Instruct	14.8	Q4	37.1 est.	Yes
DeepSeek-R1-Distill-Qwen-14B	14.8	Q4	37.1 est.	Yes
DeepSeek-V4-Flash	158.1	Q4	34.0 est.	No
Devstral-Small-2507	23.6	Q4	23.3 est.	Yes
Mistral-Small-3.2-24B-Instruct-2506	24.0	Q4	22.9 est.	Yes
Magistral-Small-2509	24.0	Q4	22.9 est.	Yes
MiniMax-M2.7	228.7	Q4	21.8 est.	No
Mistral-7B-Instruct-v0.3	7.2	FP16	21.0 est.	Yes
gemma-3-27b-it	27.4	Q4	20.2 est.	Yes
Qwen2.5-7B-Instruct	7.6	FP16	19.9 est.	Yes
Qwen2.5-Coder-7B-Instruct	7.6	FP16	19.9 est.	Yes
DeepSeek-R1-Distill-Qwen-7B	7.6	FP16	19.9 est.	Yes
GLM-4.7-Flash	31.2	FP16	19.2 est.	No
Llama-3.1-8B-Instruct	8.0	FP16	18.9 est.	Yes
Qwen3-8B	8.2	FP16	18.4 est.	Yes
DeepSeek-R1-0528-Qwen3-8B	8.2	FP16	18.4 est.	Yes
Llama-3.1-70B-Instruct	70.6	Q4	17.8 est.	No
Llama-3.3-70B-Instruct	70.6	Q4	17.8 est.	No
DeepSeek-R1-Distill-Llama-70B	70.6	Q4	17.6 est.	No
Qwen3-32B	32.8	Q4	16.8 est.	Yes
Qwen2.5-32B-Instruct	32.8	Q4	16.8 est.	Yes
Qwen2.5-Coder-32B-Instruct	32.8	Q4	16.8 est.	Yes
DeepSeek-R1-Distill-Qwen-32B	32.8	Q4	16.8 est.	Yes
Qwen2.5-72B-Instruct	72.7	Q4	16.1 est.	No
Devstral-Small-2507	23.6	FP16	15.1 est.	No
Mistral-Small-3.2-24B-Instruct-2506	24.0	FP16	14.3 est.	No
Magistral-Small-2509	24.0	FP16	14.3 est.	No
Qwen3-Coder-Next	79.7	FP16	12.8 est.	No
gemma-3-12b-it	12.2	FP16	12.6 est.	Yes
Qwen3-Next-80B-A3B-Instruct	81.3	FP16	12.3 est.	No
Phi-4	14.7	FP16	10.3 est.	Yes
Phi-4-reasoning	14.7	FP16	10.3 est.	Yes
Qwen3-14B	14.8	FP16	10.2 est.	Yes
Qwen2.5-14B-Instruct	14.8	FP16	10.2 est.	Yes
Qwen2.5-Coder-14B-Instruct	14.8	FP16	10.2 est.	Yes
DeepSeek-R1-Distill-Qwen-14B	14.8	FP16	10.2 est.	Yes
gemma-3-27b-it	27.4	FP16	9.8 est.	No
Qwen3-235B-A22B	235.1	Q4	9.1 est.	No
Qwen3-235B-A22B-Instruct-2507	235.1	Q4	9.1 est.	No
Mixtral-8x7B-Instruct-v0.1	46.7	FP16	8.2 est.	No
Qwen3-32B	32.8	FP16	5.7 est.	No
Qwen2.5-32B-Instruct	32.8	FP16	5.7 est.	No
Qwen2.5-Coder-32B-Instruct	32.8	FP16	5.7 est.	No
DeepSeek-R1-Distill-Qwen-32B	32.8	FP16	5.7 est.	No
Llama-3.1-70B-Instruct	70.6	FP16	0.6 est.	No
Llama-3.3-70B-Instruct	70.6	FP16	0.6 est.	No
DeepSeek-R1-Distill-Llama-70B	70.6	FP16	0.6 est.	No
Qwen2.5-72B-Instruct	72.7	FP16	0.5 est.	No