EdgeBench · NVIDIA DGX Spark · GB10 · 128 GB

What runs best on a DGX Spark

Open-weight models measured on one DGX Spark: how often each gets real work right, how fast it writes, and how long you wait for each right answer. Filter, sort and compare them below.

Measured so far49 of 138 configurations88 queued · updated 25 October 2026

Hardware: GB10 · 128 GB unified memory
Software: DGX OS 7.2.3 · CUDA 13.0
Engines: vLLM v0.31.0 · SGLang nightly 2026-09-09 · llama.cpp v0.6.0

Engine ·

49 of 49 measured

Fastest to a right answer12.4 sQwen3-Next-80B-A3B-Instruct · NVIDIA NVFP4per solved task (95% range 8.5–17.4 s), at 91.7% quality

Highest quality97.7%Qwen3.8-27B · RadixArk NVFP4The 95% ranges of 24 others overlap it, so they are not ranked apart

Fastest writing101.4 tok/sQwen3.6-35B-A3B · NVIDIA NVFP4 + MTP (3 tokens)after a 4k-token prompt, at 84.7% quality

Quality against time per solved task

Up and to the left is better. Seconds per solved task adds the time to read every prompt and write every reply, wrong answers included, and divides by the tasks solved. Lines show each figure’s 95% range; the Pareto front joins the configurations nothing beats on both. Click a mark to compare it.

vLLMSGLangllama.cppStandardSpeculative decodingPareto front95% range

Quality, % (higher is better)

Seconds per solved task, log scale (lower is faster)

On the Pareto front, fastest first

Qwen3-Next-80B-A3B-Instruct · NVIDIA NVFP4 12.4 s (8.5–17.4) · 91.7% (84.7–97.2%)
Ornith 1.5 35B-A3B · NVFP4 + MTP (1 token) 12.6 s (9.0–17.2) · 92.6% (87.5–97.7%)
Gemma 4 26B-A4B-IT · BF16 15.3 s (10.8–20.7) · 93.1% (88.9–97.2%)
gpt-oss-120b · OpenAI MXFP4 19.5 s (15.2–24.7) · 93.5% (88.9–97.9%)
Qwen3.8-27B · RadixArk NVFP4 + DFlash2 (8 tokens) on SGLang 42.6 s (30.1–58.9) · 94.4% (89.8–98.6%)
Gemma 4 31B-IT · NVIDIA NVFP4 48.8 s (33.4–67.3) · 97.2% (93.1–100.0%)
Qwen3.8-27B · RadixArk NVFP4 95.1 s (69.2–126.7) · 97.7% (94.0–100.0%)

Source: EdgeBench, measured on one NVIDIA DGX Spark · as of 25 October 2026

Click a column to sort. Pick up to 4 rows with to compare them.

#	Configuration
1	Qwen3.8-27B · RadixArk NVFP4vLLM· 4-bit	97.794.0–100.0	95.169.2–126.7	200	12.1	10.3	27.0	27B	2.85
2	Gemma 4 31B-IT · NVIDIA NVFP4vLLM· 4-bit	97.293.1–100.0	48.833.4–67.3	146	6.8	6.0	107.4	31B	5.29
3	Qwen3.8-27B · Qwen FP8vLLM· 8-bit	96.892.6–100.0	148.6106.2–203.0	156	7.8	7.0	30.5	27B	4.23
4	Qwen3.8-27B · NVIDIA NVFP4vLLM· 4-bit	96.391.7–100.0	92.067.3–122.3	202	12.0	10.2	27.5	27B	3.05
5	Qwen3.8-27B · RadixArk NVFP4 (BF16 lm_head) + DFlash2 (8 tokens) on SGLangSGLang· speculative decoding· 4-bit	95.891.7–99.1	48.835.4–65.6	268	28.4	26.2	32.2	27B	1.20
6	Qwen3.8-27B · RadixArk NVFP4 on SGLangSGLang· 4-bit	95.891.7–99.1	99.269.6–135.8	145	12.5	10.7	31.3	27B	2.80
7	Qwen3.8-27B · RadixArk NVFP4 (BF16 lm_head) on SGLangSGLang· 4-bit	95.891.7–99.1	118.881.6–164.6	133	10.9	9.5	31.6	27B	3.18
8	Qwen3.8-27B · RadixArk NVFP4 + DFlash2 (8 tokens) on SGLangSGLang· speculative decoding· 4-bit	94.489.8–98.6	42.630.1–58.9	272	30.6	25.2	31.2	27B	1.12
9	Qwen3.8-27B · Qwen FP8 + DFlash2 (7 tokens)vLLM· speculative decoding· 8-bit	94.088.4–98.6	79.656.4–106.8	229	18.6	15.7	31.0	27B	1.84
10	gpt-oss-120b · OpenAI MXFP4vLLM· 4-bit	93.588.9–97.9	19.515.2–24.7	467	34.7	27.3	14.2	117B · 5B active	1.04
11	Gemma 4 26B-A4B-IT · BF16vLLM· 16-bit	93.188.9–97.2	15.310.8–20.7	454	23.7	21.2	28.0	26B · 4B active	1.37
12	Qwen3.8-27B · RadixArk NVFP4 (BF16 lm_head)vLLM· 4-bit	93.187.5–97.7	129.889.8–180.4	176	10.6	9.2	27.7	27B	3.17
13	Ornith 1.5 35B-A3B · NVFP4 + MTP (1 token)vLLM· speculative decoding· 4-bit	92.687.5–97.7	12.69.0–17.2	682	92.9	71.1	9.6	35B · 3B active	0.37
14	Ornith 1.5 35B-A3B · NVFP4vLLM· 4-bit	92.186.6–97.2	15.711.2–21.3	570	77.0	55.8	9.1	35B · 3B active	0.39
15	Qwen3-Next-80B-A3B-Instruct · NVIDIA NVFP4vLLM· 4-bit	91.784.7–97.2	12.48.5–17.4	726	39.2	31.6	11.8	80B · 3B active	0.85
16	Qwen3.6-27B · Qwen FP8vLLM· 8-bit	91.486.3–96.3	321.4247.2–409.3	99	7.8	7.0	30.6	27B	4.29
17	Nemotron 3 Super 120B-A12B · NVIDIA NVFP4vLLM· 4-bit	90.785.0–95.8	69.051.6–89.6	177	15.9	15.5	19.7	120B · 12B active	2.27
18	Qwen3.6-27B · Unsloth NVFP4vLLM· 4-bit	90.585.0–95.8	223.5169.6–291.6	119	11.1	9.6	39.5	27B	3.06
19	Qwen3.6-27B · NVIDIA NVFP4vLLM· 4-bit	90.384.7–95.4	225.5170.8–290.6	114	12.0	10.2	48.3	27B	3.04
20	Step-3.7-Flash 198B-A11B · IQ4_XSllama.cpp· 4-bit	89.883.3–95.8	101.676.9–131.1	59	20.8	14.0	88.5	198B · 11B active	3.16
21	Qwen3.6-27B · rdtand PrismaQuant 5.5-bitvLLM· 5–6-bit	89.884.3–94.9	245.8184.7–323.0	119	10.4	9.0	29.8	27B	3.32
22	Qwen3.6-27B aeon ultimate uncensored · BF16vLLM· 16-bit	89.884.7–94.9	647.4496.3–828.4	55	4.5	4.2	41.1	27B	8.79
23	Ornith 1.0 35B · Q8_0llama.cpp· 8-bit	89.683.1–95.4	51.939.9–66.5	130	50.6	33.6	32.2	35B · 3B active	0.88
24	Qwen3.6-27B · Qwen FP8 + DFlash (10 tokens)vLLM· speculative decoding· 8-bit	89.484.7–94.4	144.9108.9–189.4	208	16.8	13.8	30.9	27B	2.08
25	Ornith 1.0 35B · s-batman NVFP4 GGUF (MTP)llama.cpp· 4-bit	88.482.4–94.0	46.634.2–61.7	154	60.3	37.5	29.0	35B · 3B active	0.80
26	Qwen3.6-27B aeon ultimate uncensored · BF16 + DFlash (10 tokens)vLLM· speculative decoding· 16-bit	88.483.3–93.5	279.7211.2–363.3	121	10.1	9.7	42.3	27B	3.70
27	Qwen3.6-35B-A3B · Unsloth NVFP4 + MTP (2 tokens)vLLM· speculative decoding· 4-bit	88.081.9–93.5	32.324.2–42.2	404	87.2	70.9	9.6	35B · 3B active	0.34
28	Qwen3.6-27B · kaitchup MoQ-5.0 GGUFllama.cpp· 5–6-bit	88.081.9–93.5	217.0164.0–284.3	54	11.6	8.9	99.7	27B	4.30
29	Qwen3.6-35B-A3B · Unsloth NVFP4-FastvLLM· 4-bit	87.080.1–93.3	38.628.8–50.5	313	69.3	51.6	8.7	35B · 3B active	0.43
30	Qwen3-Coder-Next 80B-A3B · Qwen FP8vLLM· 8-bit	86.879.2–93.8	13.49.1–18.7	468	40.5	33.2	14.2	80B · 3B active	0.85
31	Qwen3.6-35B-A3B · Unsloth NVFP4vLLM· 4-bit	86.879.9–93.1	41.030.7–53.3	288	68.3	51.3	8.9	35B · 3B active	0.44
32	Qwen3.6-35B-A3B · NVIDIA NVFP4vLLM· 4-bit	86.379.6–92.8	35.426.5–46.3	319	76.4	56.2	9.3	35B · 3B active	0.43
33	GLM-4.5-Air · cyankiwi AWQ 4-bitvLLM· 4-bit	86.178.7–93.1	109.677.3–150.6	83	20.3	11.4	56.8	106B · 12B active	1.85
34	Qwen3.6-27B aeon uncensored text · NVFP4 MTP-XSvLLM· 4-bit	86.179.6–92.6	263.7199.8–341.8	116	12.1	10.3	26.0	27B	2.76
35	Qwen3.6-35B-A3B · Unsloth UD-Q4_K_Mllama.cpp· 4-bit	85.278.2–91.7	49.837.7–64.6	142	60.3	37.7	30.8	35B · 3B active	0.83
36	Qwen3.5-122B-A10B · NVIDIA NVFP4vLLM· 4-bit	85.277.3–92.1	196.3150.5–250.1	92	14.3	13.2	24.6	122B · 10B active	2.48
37	Qwen3.6-35B-A3B · NVIDIA NVFP4 + MTP (3 tokens)vLLM· speculative decoding· 4-bit	84.777.3–91.7	28.521.3–37.3	451	101.4	80.7	9.8	35B · 3B active	0.33
38	Qwen3.6-27B aeon uncensored text · NVFP4 MTP-XS + MTP (3 tokens)vLLM· speculative decoding· 4-bit	84.578.2–90.7	181.1135.2–235.4	163	19.0	16.2	29.2	27B	1.77
39	Qwen3-Coder-30B-A3B-Instruct · Unsloth Q4_K_Mllama.cpp· 4-bit	84.376.4–91.4	17.79.9–27.1	167	46.1	14.5	82.2	30B · 3B active	1.56
40	GLM-4.7-Flash 30B-A3B · Q4_K_Mllama.cpp· 4-bit	84.377.6–91.0	83.953.9–123.0	38	39.2	12.2	221.0	30B · 3B active	1.14
41	qwythos-9b-claude-mythos-5-1m · BF16vLLM· 16-bit	82.673.8–90.7	125.979.7–185.9	234	12.7	11.6	12.9	9B	3.03
42	Qwen3-30B-A3B · NVIDIA NVFP4vLLM· 4-bit	82.475.0–89.4	37.828.1–49.1	348	68.4	not reached	not reached	30B · 3B active	0.49
43	Ornith 1.0 35B · s-batman NVFP4 GGUF (MTP) + MTP draft (3 tokens)llama.cpp· speculative decoding· 4-bit	81.974.1–89.4	48.235.5–64.6	179	66.1	39.6	31.0	35B · 3B active	0.66
44	Qwen-AgentWorld-35B-A3B · BF16vLLM· 16-bit	81.772.9–89.6	104.881.8–132.8	120	30.6	26.8	12.5	35B · 3B active	1.09
45	Gemma 4 12B-IT · Unsloth Q4_K_Mllama.cpp· 4-bit	79.470.4–87.7	149.0111.2–198.6	70	17.3	12.8	65.2	12B	2.89
46	Gemma 4 12B Agentic v2 (Fable5×Composer2.5) · Q4_K_Mllama.cpp· 4-bit	77.369.4–85.0	60.738.1–91.6	120	16.6	12.6	65.2	12B	2.98
47	Mellum2 12B-A2.5B Opus Thinking · Q4_K_Mllama.cpp· 4-bit	69.061.1–76.9	15.19.9–22.3	373	68.8	44.3	22.5	12B · 3B active	0.86
48	Gemma 4 12B Coder (Fable5×Composer2.5) · Q4_K_Mllama.cpp· 4-bit	54.646.1–62.7	53.437.5–75.0	134	16.9	12.6	65.1	12B	3.00
49	Gemma 4 12B Opus Reasoning · Q4_K_Mllama.cpp· 4-bit	49.541.4–57.6	185.1125.6–267.7	43	16.9	12.5	65.2	12B	2.98

Writing speed as the prompt grows

Tokens written per second once the prompt is 4k, 50k or 100k tokens long, fastest at the longest prompt first. “Not reached” means the model’s context ends before that length.

vLLMSGLangllama.cpp

Qwen3.6-35B-A3B · NVIDIA + MTP (3 tokens)4k101.450k89.6100k80.7
Ornith 1.5 35B-A3B + MTP (1 token)4k92.950k76.7100k71.1
Qwen3.6-35B-A3B · Unsloth + MTP (2 tokens)4k87.250k79.4100k70.9
Qwen3.6-35B-A3B · NVIDIA4k76.450k64.4100k56.2
Ornith 1.5 35B-A3B4k77.050k64.5100k55.8
Qwen3.6-35B-A3B · Unsloth NVFP4-Fast4k69.350k59.1100k51.6
Qwen3.6-35B-A3B · Unsloth NVFP44k68.350k58.3100k51.3
Mellum2 12B-A2.5B Opus Thinking4k68.850k44.1100k44.3
Ornith 1.0 35B · s-batman (MTP) + MTP draft (3 tokens)4k66.150k39.0100k39.6
Qwen3.6-35B-A3B · Unsloth UD-Q4_K_M4k60.350k37.4100k37.7
Ornith 1.0 35B · s-batman (MTP)4k60.350k37.2100k37.5
Ornith 1.0 35B4k50.650k33.4100k33.6
Qwen3-Coder-Next 80B-A3B · Qwen4k40.550k36.4100k33.2
Qwen3-Next-80B-A3B-Instruct · NVIDIA4k39.250k34.9100k31.6
gpt-oss-120b · OpenAI4k34.750k30.7100k27.3
Qwen-AgentWorld-35B-A3B4k30.650k28.5100k26.8
Qwen3.8-27B · RadixArk (BF16 lm_head) + DFlash2 (8 tokens) on SGLang4k28.450k25.2100k26.2
Qwen3.8-27B · RadixArk + DFlash2 (8 tokens) on SGLang4k30.650k32.3100k25.2
Gemma 4 26B-A4B-IT4k23.750k22.2100k21.2
Qwen3.6-27B aeon uncensored text · MTP-XS + MTP (3 tokens)4k19.050k17.7100k16.2
Qwen3.8-27B · Qwen + DFlash2 (7 tokens)4k18.650k18.4100k15.7
Nemotron 3 Super 120B-A12B · NVIDIA4k15.950k15.7100k15.5
Qwen3-Coder-30B-A3B-Instruct · Unsloth4k46.150k12.7100k14.5
Step-3.7-Flash 198B-A11B4k20.850k13.1100k14.0
Qwen3.6-27B · Qwen + DFlash (10 tokens)4k16.850k17.8100k13.8
Qwen3.5-122B-A10B · NVIDIA4k14.350k13.7100k13.2
Gemma 4 12B-IT · Unsloth4k17.350k12.7100k12.8
Gemma 4 12B Agentic v2 (Fable5×Composer2.5)4k16.650k12.5100k12.6
Gemma 4 12B Coder (Fable5×Composer2.5)4k16.950k12.5100k12.6
Gemma 4 12B Opus Reasoning4k16.950k12.5100k12.5
GLM-4.7-Flash 30B-A3B4k39.250k10.8100k12.2
qwythos-9b-claude-mythos-5-1m4k12.750k12.1100k11.6
GLM-4.5-Air · cyankiwi4k20.350k14.7100k11.4
Qwen3.8-27B · RadixArk on SGLang4k12.550k11.6100k10.7
Qwen3.6-27B aeon uncensored text · MTP-XS4k12.150k11.1100k10.3
Qwen3.8-27B · RadixArk4k12.150k11.1100k10.3
Qwen3.8-27B · NVIDIA4k12.050k11.0100k10.2
Qwen3.6-27B · NVIDIA4k12.050k11.0100k10.2
Qwen3.6-27B aeon ultimate uncensored + DFlash (10 tokens)4k10.150k10.2100k9.7
Qwen3.6-27B · Unsloth4k11.150k10.3100k9.6
Qwen3.8-27B · RadixArk (BF16 lm_head) on SGLang4k10.950k10.2100k9.5
Qwen3.8-27B · RadixArk (BF16 lm_head)4k10.650k9.8100k9.2
Qwen3.6-27B · rdtand4k10.450k9.7100k9.0
Qwen3.6-27B · kaitchup MoQ-5.04k11.650k8.9100k8.9
Qwen3.8-27B · Qwen4k7.850k7.4100k7.0
Qwen3.6-27B · Qwen4k7.850k7.4100k7.0
Gemma 4 31B-IT · NVIDIA4k6.850k6.3100k6.0
Qwen3.6-27B aeon ultimate uncensored4k4.550k4.3100k4.2
Qwen3-30B-A3B · NVIDIA4k68.450knot reached100knot reached

Source: EdgeBench, measured on one NVIDIA DGX Spark · as of 25 October 2026

A small draft model guesses several tokens ahead and the main model checks them in one pass, keeping the ones it would have written itself. It should change speed, not answers, so each row is compared task by task with the same model on the same engine without drafting; a paired test with p below 0.05 would count as answers changed.

With speculative decoding	Writing speed, tok/s at 4k	Speed-up	Seconds per solved task	Answers
Qwen3.8-27B · RadixArk NVFP4 (BF16 lm_head) + DFlash2 (8 tokens) on SGLangAgainst Qwen3.8-27B · RadixArk NVFP4 (BF16 lm_head) on SGLang	10.9 → 28.4	2.61×	118.8 → 48.895% ranges 81.6–164.6 and 35.4–65.6	No change detectedp = 1.00 across 86 tasks
Qwen3.8-27B · RadixArk NVFP4 + DFlash2 (8 tokens) on SGLangAgainst Qwen3.8-27B · RadixArk NVFP4 on SGLang	12.5 → 30.6	2.44×	99.2 → 42.695% ranges 69.6–135.8 and 30.1–58.9	No change detectedp = 0.45 across 86 tasks
Qwen3.8-27B · Qwen FP8 + DFlash2 (7 tokens)Against Qwen3.8-27B · Qwen FP8	7.8 → 18.6	2.37×	148.6 → 79.695% ranges 106.2–203.0 and 56.4–106.8	No change detectedp = 0.34 across 86 tasks
Qwen3.6-27B aeon ultimate uncensored · BF16 + DFlash (10 tokens)Against Qwen3.6-27B aeon ultimate uncensored · BF16	4.5 → 10.1	2.27×	647.4 → 279.795% ranges 496.3–828.4 and 211.2–363.3	No change detectedp = 0.45 across 86 tasks
Qwen3.6-27B · Qwen FP8 + DFlash (10 tokens)Against Qwen3.6-27B · Qwen FP8	7.8 → 16.8	2.15×	321.4 → 144.995% ranges 247.2–409.3 and 108.9–189.4	No change detectedp = 0.58 across 86 tasks
Qwen3.6-27B aeon uncensored text · NVFP4 MTP-XS + MTP (3 tokens)Against Qwen3.6-27B aeon uncensored text · NVFP4 MTP-XS	12.1 → 19.0	1.56×	263.7 → 181.195% ranges 199.8–341.8 and 135.2–235.4	No change detectedp = 0.33 across 86 tasks
Qwen3.6-35B-A3B · NVIDIA NVFP4 + MTP (3 tokens)Against Qwen3.6-35B-A3B · NVIDIA NVFP4	76.4 → 101.4	1.33×	35.4 → 28.595% ranges 26.5–46.3 and 21.3–37.3	No change detectedp = 0.58 across 86 tasks
Qwen3.6-35B-A3B · Unsloth NVFP4 + MTP (2 tokens)Against Qwen3.6-35B-A3B · Unsloth NVFP4	68.3 → 87.2	1.28×	41.0 → 32.395% ranges 30.7–53.3 and 24.2–42.2	No change detectedp = 0.71 across 86 tasks
Ornith 1.5 35B-A3B · NVFP4 + MTP (1 token)Against Ornith 1.5 35B-A3B · NVFP4	77.0 → 92.9	1.21×	15.7 → 12.695% ranges 11.2–21.3 and 9.0–17.2	No change detectedp = 1.00 across 86 tasks
Ornith 1.0 35B · s-batman NVFP4 GGUF (MTP) + MTP draft (3 tokens)Against Ornith 1.0 35B · s-batman NVFP4 GGUF (MTP)	60.3 → 66.1	1.10×	46.6 → 48.295% ranges 34.2–61.7 and 35.5–64.6	No change detectedp = 0.06 across 86 tasks

Measure

Speed, split

Reading the prompt (prefill) and writing the answer (decode) are timed apart, at exact prompt lengths counted in the model’s own tokens and built from real source code.

Measure

Quality, graded on the box

Tool calls, code, maths, instructions, long-context lookups and bug repairs are graded by running the code or matching the exact answer, with no model as judge, on the exact compressed file being served.

Measure

Time per solved task

Seconds per solved task adds the time to read every prompt and write every reply, wrong answers included, and divides by the tasks solved, so a fast model that rambles or answers wrongly does not win.

Measure

Speculative decoding, checked

When a small draft model guesses ahead and the main model checks its guesses, the answers are compared task by task with the same model running without the draft, to show whether drafting changed them.

One serving policy for every model

Standard rows run every model with the same settings: an 8-bit cache of the prompt’s working state, reuse of repeated prompt openings, up to 16 requests at once and the same share of memory. Differences come from the model and its compressed weights, not from tuning. Speculative rows add a drafter to the same weights and change nothing else. Speeds are medians of repeated runs at each prompt length.

What each measure means

Quality.: The share of tasks solved, averaged over six task families. Two scores differ only when their 95% ranges do not overlap.
Time per solved task.: Seconds for one user alone: reading every prompt and writing every reply at the measured speeds, wrong answers included, divided by the tasks solved.
Solved per hour.: Tasks solved in an hour with up to 16 requests running at once.
Writing speed.: Tokens written per second (decode) after a 4k- or 100k-token prompt, the median of repeated runs.
First token.: Seconds before the reply starts for a new 50k-token prompt, with nothing cached. “Not reached” means the model’s context ends sooner.
Usable depth.: The longest prompt at which it still writes at least 10 tokens a second, starts replying within 60 seconds and finds a planted note every time.
Size.: Total parameters; a mixture of experts uses only the active ones for each token.
Energy.: Joules per written token after a 4k-token prompt, from the GPU’s own power reading.

Limits: read before quoting a number

One DGX Spark and one run per configuration. The box’s temperature is recorded, not controlled.
Energy is the GPU’s own power reading, not power drawn at the wall.
Quality is a compact check of 74–86 tasks per configuration, so its ranges are wide. It is not a full coding benchmark.
Each engine runs one pinned version; other versions may be faster or slower.
Queued configurations are not results until they are measured.

88 configurations are waiting their turn on the DGX Spark. They join the views above once measured; until then they are not results.

vLLM56 queued

Apertus 70B Instruct · Red Hat FP8
Apriel 1.6 15B Thinker · BF16
DeepSeek-R1-Distill-Qwen-32B · Red Hat W4A16
Devstral Small 2 24B · Mistral FP8
DiffusionGemma 26B-A4B · NVIDIA NVFP4
ERNIE 4.5 21B-A3B Thinking · cyankiwi AWQ 4-bit
EXAONE 4.5 33B · LG FP8
Gemma 4 E4B-IT · Google QAT W4A16
gpt-oss-20b · OpenAI MXFP4
Granite 4.2 30B · IBM NVFP4
Hermes 4 14B · Nous FP8
Jamba2 Mini · AI21 FP8
K2 Horizon MoVA 36B-A4B · BF16
KAT-Coder V2.5 Dev · cyankiwi AWQ 4-bit
Laguna S 2.1 118B-A8B · NVFP4
Laguna S 2.1 118B-A8B · NVFP4 + DFlash (3 tokens)
Laguna XS 2.1 33B-A3B · BF16
Laguna XS 2.1 33B-A3B · BF16 + DFlash (15 tokens)
LFM2 24B-A2B · BF16
LFM2.5 8B-A1B · cyankiwi AWQ 4-bit
Ling 3.0 Flash · inclusionAI INT4
Ling 3.0 Tiny · inclusionAI FP8
Llama 3.1 8B Instruct · NVIDIA NVFP4
Llama 3.3 70B Instruct · NVIDIA NVFP4
Llama 3.3 Nemotron Super 49B v1.5 · NVIDIA NVFP4
MiMo V2.6 Distill Qwen 9B · cyankiwi AWQ 4-bit
MiniCPM5 2B · OpenBMB 4-bit
Mistral Small 3.2 24B · BF16
Mistral Small 4 119B · Mistral NVFP4
Muse Glimmer 30B · NVIDIA NVFP4
Nanbeige 4.2 3B · BF16
Nemotron 3 Nano 30B-A3B · NVIDIA NVFP4
Nemotron 3 Nano 4B · NVIDIA FP8
Nemotron 3 Nano Omni 30B-A3B Reasoning · NVIDIA NVFP4
Nemotron 3.5 Lightning 30B-A3B · NVIDIA NVFP4
Nemotron Cascade 2 30B-A3B · cyankiwi AWQ 4-bit
Nemotron Labs 3 Puzzle 75B-A9B · NVIDIA NVFP4
Nemotron Nano 9B v2 · NVIDIA NVFP4
North Mini Code 1.0 · Cohere FP8
OLMo 3.1 32B Instruct · BF16
Ornith 1.5 9B · Ornith NVFP4
Phi-4 · Red Hat W4A16
Phi-4 reasoning plus · NVIDIA NVFP4
Qwen2.5-VL-7B · NVIDIA NVFP4
Qwen3-14B · NVIDIA NVFP4
Qwen3-32B · NVIDIA NVFP4
Qwen3-8B · NVIDIA NVFP4
Qwen3-VL-30B-A3B-Instruct · Qwen FP8
Qwen3-VL-8B-Instruct · Qwen FP8
Qwen3.5-35B-A3B · Qwen FP8
Qwen3.5-9B · Red Hat W4A16
Qwen3.8-27B · RadixArk NVFP4 + DSpark (7 tokens)
Qwopus 3.6 27B Coder · FP8
Sarvam 30B · Sarvam FP8
Seed-OSS 36B Instruct · AngelSlim FP8
Step3-VL 10B · StepFun FP8

SGLang10 queued

Gemma 4 26B-A4B-IT · BF16 on SGLang
gpt-oss-120b · OpenAI MXFP4 on SGLang
gpt-oss-20b · OpenAI MXFP4 on SGLang
Llama 3.3 70B Instruct · NVIDIA NVFP4 on SGLang
Nemotron 3 Nano 30B-A3B · NVIDIA NVFP4 on SGLang
Qwen3-30B-A3B · NVIDIA NVFP4 on SGLang
Qwen3-Next-80B-A3B-Instruct · NVIDIA NVFP4 on SGLang
Qwen3.6-35B-A3B · NVIDIA NVFP4 on SGLang
Qwen3.8-27B · NVIDIA NVFP4 on SGLang
Qwen3.8-27B · Qwen FP8 on SGLang

llama.cpp11 queued

Command A+ · bartowski Q3_K_XL
GLM-5.3-Flash · Unsloth UD-IQ2_XXS
Hy3 · bartowski Q2_K_L
Kimi Linear 48B-A3B · bartowski GGUF
MiMo V2.6 Flash · AesSedai GGUF (2.5-bit)
MiniMax-M2.7 · Unsloth UD-Q3_K_XL
Phi-4 reasoning vision 15B · DevQuasar Q8_0
Qwen3.8-Flash-Next · ggml-org IQ4_NL
Qwen3.8-Flash-Next · ISTA-DASLab GSQ-RCO IQ2_XS
Ternary Bonsai 27B · PrismML Q2_g64
Ternary Bonsai 8B · PrismML Q2_0_g64

llama.cpp (PrismML fork)5 queued

Bonsai 27B · PrismML 1-bit Q1_0
Bonsai 8B · PrismML 1-bit Q1_0
Ternary Bonsai 2 27B · PrismML PQ2_0
Ternary Bonsai 2 27B · PrismML PTQ1_0
Ternary Bonsai 27B · PrismML PQ2_0

Strata1 queued

Qwen3.8-Flash-Next · ISTA-DASLab GSQ-RCO IQ2_XS on Strata

FreeToken5 queued

Gemma 4 26B-A4B-IT · BF16 on FreeToken
gpt-oss-120b · OpenAI MXFP4 on FreeToken
gpt-oss-20b · OpenAI MXFP4 on FreeToken
Qwen3.6-35B-A3B · NVIDIA NVFP4 on FreeToken
Qwen3.8-27B · RadixArk NVFP4 on FreeToken

Also waiting for access: Qwen3-Coder-Next 80B-A3B · saricles NVFP4. Its weights are gated on Hugging Face, and it runs once an access token is provided.

What runs best on a DGX Spark

Quality against time per solved task

On the Pareto front, fastest first

Quality

Time per solved task

Writing speed at 4k

First token at 50k

Solved per hour at load

Energy at 4k

Writing speed as the prompt grows

Writing speed as the prompt grows

Speed, split

Quality, graded on the box

Time per solved task

Speculative decoding, checked

One serving policy for every model

What each measure means

Limits: read before quoting a number