New Solar-Open2-250B Release. (2nd K-LLM after Motif-3 Beta)

Accelerated Computing DGX Spark / GB10 User Forum DGX Spark / GB10
1

Solar-Open2-250B (250B-A15B) is released together with INT4 and NVFP4 quantization.

Just for information, South Korean government is sponsoring a national AI initiative aimed at achieving technological sovereignty from 5 different companies. Motif-3 is also one of them. Enjoy the competition.

image
image2932×1763 383 KB

2

got the int4 version up on 2 spark. there is no MTP instruction so did not tried with MTP.

model test t/s peak t/s ttfr (ms) est_ppt (ms) e2e_ttft (ms)
solar-open2 pp2048 2227.11 ± 91.97 1003.69 ± 25.03 802.05 ± 25.03 1003.69 ± 25.03
solar-open2 tg32 14.88 ± 0.09 15.33 ± 0.47
solar-open2 pp2048 @ d4096 2062.98 ± 3.14 2874.55 ± 85.02 2672.91 ± 85.02 2874.55 ± 85.02
solar-open2 tg32 @ d4096 14.92 ± 0.08 16.00 ± 0.00
solar-open2 pp2048 @ d8192 2011.85 ± 13.35 4792.42 ± 73.81 4590.77 ± 73.81 4792.42 ± 73.81
solar-open2 tg32 @ d8192 14.83 ± 0.09 15.33 ± 0.47
solar-open2 pp2048 @ d16384 1980.75 ± 31.81 8625.29 ± 87.29 8423.65 ± 87.29 8638.41 ± 87.87
solar-open2 tg32 @ d16384 14.74 ± 0.19 15.67 ± 0.47
solar-open2 pp2048 @ d32768 1875.92 ± 20.07 16933.18 ± 182.91 16731.54 ± 182.91 16954.82 ± 183.35
solar-open2 tg32 @ d32768 14.29 ± 0.41 15.33 ± 0.47
3

Is it intelligent or benchmaxxed?

4

I experimented with the NVFP4 version, output on my workflows was close but subjectively not quite as good as mimo-v2.5, so the benchmarks presented are likely accurate.

Solar Open2 250B (NVFP4) on a 2× GB10 pair — FP8 KV cache A/B on a hybrid linear-attention MoE - DGX Spark / GB10 User Forum / DGX Spark / GB10 - NVIDIA Developer Forums