SnowLLM / Benchmark

Benchmark

Measured performance on Ryzen AI Max+ 395. Compare engines, models and context lengths with the full results behind every chart.

Benchmarks

Measured on Ryzen AI Max+ 395. Methodology ↗

All measured results

Complete measurements on Ryzen AI Max+ 395. Decode and prefill rates are in tokens per second; TTFT is in seconds and step time is in milliseconds. K = 1024 tokens.

DeepSeek-V4-Flash IQ2_XXS
Engine	Variant	Target input (tokens)	Actual input (tokens)	decode tok/s	prefill tok/s	TTFT s	ms/step	AL
SnowLLM	DSpark, ctx 132096	1,024	1,033	23.09	360.6	2.865	115.25	2.679
SnowLLM	DSpark, ctx 132096	8,192	8,202	21.9	629.8	13.024	115.37	2.567
SnowLLM	DSpark, ctx 132096	32,768	32,777	20.78	632.1	51.853	119.08	2.47
SnowLLM	DSpark, ctx 132096	131,072	131,082	19.93	681.4	192.38	131.25	2.615
SnowLLM	DSpark, ctx 227328	1,024	1,033	23.06	360.7	2.864	115.54	2.679
SnowLLM	DSpark, ctx 227328	8,192	8,202	21.9	629.1	13.037	115.51	2.551
SnowLLM	DSpark, ctx 227328	32,768	32,777	20.73	632.5	51.823	119.37	2.47
SnowLLM	DSpark, ctx 227328	131,072	131,082	18.43	680.4	192.643	132.81	2.42
SnowLLM	DSpark, ctx 227328	225,280	225,287	16.66	513.4	438.848	144.31	2.408
SnowLLM	no spec, ctx 132096	1,024	1,033	18.15	370.4	2.789	55.08	1.0
SnowLLM	no spec, ctx 132096	8,192	8,202	18.13	634.4	12.929	55.15	1.0
SnowLLM	no spec, ctx 132096	32,768	32,777	17.78	636.0	51.536	56.24	1.0
SnowLLM	no spec, ctx 132096	131,072	131,082	16.56	682.5	192.055	60.39	1.0
llama.cpp	no spec, ctx 132096	1,024	1,033	15.36	179.2	5.766	65.09	1.0
llama.cpp	no spec, ctx 132096	8,192	8,202	12.91	177.1	46.311	77.47	1.0
llama.cpp	no spec, ctx 132096	32,768	32,777	11.75	149.4	219.439	85.12	1.0
llama.cpp	no spec, ctx 132096	131,072	131,082	8.77	72.7	1802.114	113.97	1.0

GLM-5.3-Flash UD-IQ2_XXS
Engine	Variant	Target input (tokens)	Actual input (tokens)	decode tok/s	prefill tok/s	TTFT s	ms/step	AL
SnowLLM	MTP k=2	1,024	1,042	27.65	437.1	2.384	86.26	2.378
SnowLLM	MTP k=2	8,192	8,211	25.76	517.1	15.878	88.11	2.284
SnowLLM	MTP k=2	32,768	32,787	26.17	507.7	64.579	88.91	2.315
SnowLLM	MTP k=2	131,072	131,091	26.92	491.1	266.947	91.64	2.487
SnowLLM	MTP k=2	225,280	225,297	26.79	480.7	468.715	93.79	2.49
llama.cpp	MTP k=2	1,024	1,042	18.91	185.6	5.615	122.38	2.314
llama.cpp	MTP k=2	8,192	8,211	17.07	208.6	39.354	132.01	2.254
llama.cpp	MTP k=2	32,768	32,787	13.78	126.5	259.255	167.34	2.305
llama.cpp	MTP k=2	131,072	131,091	7.42	48.6	2694.976	313.83	2.329
llama.cpp	MTP k=2	225,280	225,297	6.17	30.7	7340.729	405.87	2.504

Qwen3.6-35B-A3B FP8
Engine	Variant	Target input (tokens)	Actual input (tokens)	decode tok/s	prefill tok/s	TTFT s	ms/step	AL
SnowLLM	DFlash	1,024	1,041	68.53	2133.8	0.488	45.66	3.12
SnowLLM	DFlash	8,192	8,209	62.53	3605.5	2.277	46.88	2.931
SnowLLM	DFlash	32,768	32,783	58.09	3019.6	10.857	49.64	2.872
SnowLLM	DFlash	131,072	131,086	42.23	1811.4	72.368	68.57	2.809
SnowLLM	DFlash	225,280	225,297	25.62	1303.4	172.852	98.49	2.476
SnowLLM	MTP k=2	1,024	1,041	67.39	2563.4	0.406	36.95	2.504
SnowLLM	MTP k=2	8,192	8,209	65.1	3851.4	2.131	37.85	2.481
SnowLLM	MTP k=2	32,768	32,783	61.9	3033.6	10.807	40.67	2.507
SnowLLM	MTP k=2	131,072	131,086	49.59	1720.8	76.177	52.55	2.573
SnowLLM	MTP k=2	225,280	225,297	39.38	1214.6	185.493	63.28	2.477

Qwen3.6-35B-A3B UD-Q4_K_XL
Engine	Variant	Target input (tokens)	Actual input (tokens)	decode tok/s	prefill tok/s	TTFT s	ms/step	AL
SnowLLM	DFlash	1,024	1,041	86.37	2131.9	0.488	37.7	3.231
SnowLLM	DFlash	8,192	8,209	78.18	3715.2	2.21	38.94	2.973
SnowLLM	DFlash	32,768	32,783	70.61	3861.2	8.49	41.94	2.982
SnowLLM	DFlash	131,072	131,086	55.01	1835.3	71.425	53.34	2.925
SnowLLM	DFlash	225,280	225,297	30.83	1311.6	171.775	83.09	2.557
llama.cpp	DFlash	1,024	1,041	57.4	874.6	1.19	54.47	3.127
llama.cpp	DFlash	8,192	8,209	45.96	1063.7	7.718	61.01	2.804
llama.cpp	DFlash	32,768	32,783	41.82	984.4	33.302	63.3	2.647
llama.cpp	DFlash	131,072	131,086	27.86	485.8	269.858	93.92	2.617
llama.cpp	DFlash	225,280	225,297	20.42	297.7	756.894	119.06	2.431
SnowLLM	MTP k=2	1,024	1,041	86.4	2145.4	0.485	28.63	2.489
SnowLLM	MTP k=2	8,192	8,209	83.84	3806.5	2.157	29.52	2.468
SnowLLM	MTP k=2	32,768	32,783	75.55	3864.4	8.483	32.43	2.452
SnowLLM	MTP k=2	131,072	131,086	54.4	1765.3	74.258	46.73	2.557
SnowLLM	MTP k=2	225,280	225,297	33.59	1244.0	181.111	72.33	2.465
SnowLLM	no spec	1,024	1,041	62.02	2245.2	0.464	16.12	1.0
SnowLLM	no spec	8,192	8,209	59.06	3936.0	2.086	16.93	1.0
SnowLLM	no spec	32,768	32,783	52.45	4052.9	8.089	19.06	1.0
SnowLLM	no spec	131,072	131,086	35.14	1764.3	74.3	28.46	1.0
SnowLLM	no spec	225,280	225,297	20.56	1267.8	177.711	48.64	1.0

Qwen3.8-27B UD-Q4_K_XL
Engine	Variant	Target input (tokens)	Actual input (tokens)	decode tok/s	prefill tok/s	TTFT s	ms/step	AL
SnowLLM	DFlash2	1,024	1,083	28.9	511.9	2.116	115.45	3.417
SnowLLM	DFlash2	8,192	8,251	24.61	602.7	13.69	117.98	2.902
SnowLLM	DFlash2	32,768	32,825	18.66	502.3	65.355	165.81	3.173
SnowLLM	DFlash2	131,072	131,128	13.84	350.5	374.126	221.68	3.154
SnowLLM	DFlash2	225,280	225,339	11.46	272.3	827.468	261.16	3.058
SnowLLM	MTP k=2	1,024	1,083	20.28	503.8	2.15	112.29	2.284
SnowLLM	MTP k=2	8,192	8,251	19.26	602.9	13.685	115.08	2.218
SnowLLM	MTP k=2	32,768	32,825	15.29	499.6	65.702	142.59	2.189
SnowLLM	MTP k=2	131,072	131,128	10.31	343.9	381.3	220.89	2.279
SnowLLM	MTP k=2	225,280	225,339	9.23	264.9	850.649	267.74	2.464
llama.cpp	MTP k=2	1,024	1,083	17.65	301.0	3.598	125.03	2.207
llama.cpp	MTP k=2	8,192	8,251	23.54	300.6	27.446	126.31	2.974
llama.cpp	MTP k=2	32,768	32,825	17.27	275.7	119.075	154.77	2.673
llama.cpp	MTP k=2	131,072	131,128	14.35	117.4	1116.969	187.05	2.684
llama.cpp	MTP k=2	225,280	225,339	9.56	66.4	3392.35	252.95	2.419

Qwen3.8-Flash-Next UD-Q3_K_XL
Engine	Variant	Target input (tokens)	Actual input (tokens)	decode tok/s	prefill tok/s	TTFT s	ms/step	AL
SnowLLM	MTP k=2, Engine	1,024	1,024	43.29	1324.3	0.773	55.01	2.391
SnowLLM	MTP k=2, Engine	8,192	8,192	42.93	1812.2	4.521	56.03	2.429
SnowLLM	MTP k=2, Engine	32,768	32,768	41.81	1842.0	17.789	56.47	2.361
SnowLLM	MTP k=2, Engine	131,072	131,072	39.84	1775.1	73.84	57.93	2.304
SnowLLM	MTP k=2, Engine	225,280	225,280	38.56	1723.8	130.686	59.12	2.284
SnowLLM	MTP k=2, HTTP	1,024	1,083	39.13	1152.0	0.94	56.33	2.208
SnowLLM	MTP k=2, HTTP	8,192	8,251	39.09	1668.7	4.944	57.49	2.239
SnowLLM	MTP k=2, HTTP	32,768	32,825	36.13	1703.3	19.272	57.87	2.118
SnowLLM	MTP k=2, HTTP	131,072	131,128	35.57	1672.7	78.392	59.74	2.171
SnowLLM	MTP k=2, HTTP	225,280	225,339	35.61	1622.5	138.881	60.92	2.191
SnowLLM	no spec, Engine	1,024	1,024	28.3	1380.0	0.742	35.69	1.0
SnowLLM	no spec, Engine	8,192	8,192	28.0	1874.5	4.37	35.75	1.0
SnowLLM	no spec, Engine	32,768	32,768	27.86	1890.8	17.331	35.85	1.0
SnowLLM	no spec, Engine	131,072	131,072	27.25	1811.9	72.338	36.67	1.0
SnowLLM	no spec, Engine	225,280	225,280	26.46	1761.8	127.87	37.76	1.0

llama.cpp AL is estimated as 1 + maximum draft width × acceptance rate; AL and derived step time are upper bounds if fewer tokens are proposed.