Spark recipe
Models with a native DGX Spark recipe / support (Spark recipe tag).
← all tag kinds
Spark recipe (78)
| Configuration | Engine | Quant | Ctx | Conc | Decode tok/s | Status | Completed |
|---|---|---|---|---|---|---|---|
| Qwen3.8-27B · SGLang · NVFP4 + DSpark · conc-32 | SGLang + DSpark | NVFP4 | 262144 | 32 | — | blocked | 2026-08-20 05:25 +0800 |
| Qwen3.8-27B · SGLang · NVFP4 + DSpark · conc-16 | SGLang + DSpark | NVFP4 | 262144 | 16 | — | blocked | 2026-08-20 05:25 +0800 |
| Qwen3.8-27B · SGLang · NVFP4 + DFlash2 · conc-8 | SGLang + DFlash2 | NVFP4 | 262144 | 8 | — | blocked | 2026-08-20 05:25 +0800 |
| Qwen3.8-27B · SGLang · NVFP4 + DFlash2 · conc-32 | SGLang + DFlash2 | NVFP4 | 262144 | 32 | — | blocked | 2026-08-20 05:25 +0800 |
| Qwen3.8-27B · SGLang · NVFP4 + DFlash2 · conc-16 | SGLang + DFlash2 | NVFP4 | 262144 | 16 | — | blocked | 2026-08-20 05:25 +0800 |
| Qwen3.8-27B · SGLang · NVFP4 + DFlash2 · conc-4 | SGLang + DFlash2 | NVFP4 | 262144 | 4 | 85.1 | done | 2026-08-20 05:24 +0800 |
| Qwen3.8-27B · SGLang · NVFP4 + DFlash2 · conc-2 | SGLang + DFlash2 | NVFP4 | 262144 | 2 | 56.2 | done | 2026-08-20 05:16 +0800 |
| Qwen3.8-27B · SGLang · NVFP4 + DFlash2 · conc-1 | SGLang + DFlash2 | NVFP4 | 262144 | 1 | 30.0 | done | 2026-08-20 05:04 +0800 |
| Qwen3.8-27B · SGLang · NVFP4 + DSpark · conc-8 | SGLang + DSpark | NVFP4 | 262144 | 8 | 118.3 | done | 2026-08-20 04:46 +0800 |
| Qwen3.8-27B · SGLang · NVFP4 + DSpark · conc-4 | SGLang + DSpark | NVFP4 | 262144 | 4 | 75.5 | done | 2026-08-20 04:17 +0800 |
| Qwen3.8-27B · SGLang · NVFP4 + DSpark · conc-2 | SGLang + DSpark | NVFP4 | 262144 | 2 | 43.4 | done | 2026-08-20 03:46 +0800 |
| Qwen3.8-27B · SGLang · NVFP4 + DSpark · conc-1 | SGLang + DSpark | NVFP4 | 262144 | 1 | 23.4 | done | 2026-08-20 03:31 +0800 |
| Nemotron-3 Puzzle 75B-A9B · vLLM · NVFP4 · conc 32 | vLLM | NVFP4 | 65536 | 32 | 177.8 | done | 2026-07-11 17:21 +08 |
| Nemotron-3 Puzzle 75B-A9B · vLLM · NVFP4 · conc 8 | vLLM | NVFP4 | 65536 | 8 | 94.0 | done | 2026-07-11 16:54 +08 |
| Nemotron-3 Puzzle 75B-A9B · vLLM · NVFP4 · conc 1 | vLLM | NVFP4 | 65536 | 1 | 19.8 | done | 2026-07-11 16:29 +08 |
| Qwen3.6-35B-A3B · vLLM-ultimate (AEON) · NVFP4 + DFlash · conc 128 (memory ceiling — watchdog trip) | vLLM (aeon-vllm-ultimate custom container, v0.23.0+aeon.sm121a.dflash) + DFlash (z-lab/Qwen3.6-35B-A3B-DFlash @31977fbe small-page rev, num_speculative_tokens 11) | NVFP4 (modelopt_mixed — W4A16_NVFP4 experts + FP8 GDN gates) | 65536 | 128 | 425.7 | done | 2026-07-04 14:17 +0800 |
| Qwen3.6-35B-A3B · vLLM-ultimate (AEON) · NVFP4 + DFlash · conc 64 | vLLM (aeon-vllm-ultimate custom container, v0.23.0+aeon.sm121a.dflash) + DFlash (z-lab/Qwen3.6-35B-A3B-DFlash @31977fbe small-page rev, num_speculative_tokens 11) | NVFP4 (modelopt_mixed — W4A16_NVFP4 experts + FP8 GDN gates) | 65536 | 64 | 420.4 | done | 2026-07-04 14:11 +0800 |
| Qwen3.6-35B-A3B · vLLM · NVFP4 · conc 128 | vLLM | NVFP4 | 65536 | 128 | 675.9 | done | 2026-07-04 13:50 +0800 |
| Qwen3.6-35B-A3B · vLLM · NVFP4 · conc 64 | vLLM | NVFP4 | 65536 | 64 | 547.9 | done | 2026-07-04 13:41 +0800 |
| Qwen3.6-35B-A3B · vLLM · NVFP4 + MTP · conc 128 (the knee) | vLLM + MTP | NVFP4 | 65536 | 128 | 749.9 | done | 2026-07-04 12:31 +0800 |
| Qwen3.6-35B-A3B · vLLM · NVFP4 + MTP · conc 64 | vLLM + MTP | NVFP4 | 65536 | 64 | 670.8 | done | 2026-07-04 12:09 +0800 |
| gpt-oss-20b · vLLM · MXFP4 · conc 8 | vLLM | MXFP4 | 65536 | 8 | 212.5 | done | 2026-07-02 07:51 +0800 |
| gpt-oss-20b · vLLM · MXFP4 · conc 1 | vLLM | MXFP4 | 65536 | 1 | 45.6 | done | 2026-07-02 07:51 +0800 |
| Qwen3.6-35B-A3B · vLLM-ultimate (AEON) · NVFP4 + DFlash · conc 1 | vLLM (aeon-vllm-ultimate custom container, v0.23.0+aeon.sm121a.dflash) + DFlash (z-lab/Qwen3.6-35B-A3B-DFlash @31977fbe small-page rev, num_speculative_tokens 11) | NVFP4 (modelopt_mixed — W4A16_NVFP4 experts + FP8 GDN gates) | 65536 | 1 | 99.8 | done | 2026-07-02 07:21 +0800 |
| Qwen3.6-35B-A3B · vLLM-ultimate (AEON) · NVFP4 + DFlash · conc 8 | vLLM (aeon-vllm-ultimate custom container, v0.23.0+aeon.sm121a.dflash) + DFlash (z-lab/Qwen3.6-35B-A3B-DFlash @31977fbe small-page rev, num_speculative_tokens 11) | NVFP4 (modelopt_mixed — W4A16_NVFP4 experts + FP8 GDN gates) | 65536 | 8 | 267.4 | done | 2026-07-02 07:21 +0800 |
| Qwen3.6-35B-A3B · vLLM-ultimate (AEON) · NVFP4 + DFlash · conc 4 | vLLM (aeon-vllm-ultimate custom container, v0.23.0+aeon.sm121a.dflash) + DFlash (z-lab/Qwen3.6-35B-A3B-DFlash @31977fbe small-page rev, num_speculative_tokens 11) | NVFP4 (modelopt_mixed — W4A16_NVFP4 experts + FP8 GDN gates) | 65536 | 4 | 210.7 | done | 2026-07-02 07:21 +0800 |
| Qwen3.6-35B-A3B · vLLM-ultimate (AEON) · NVFP4 + DFlash · conc 32 | vLLM (aeon-vllm-ultimate custom container, v0.23.0+aeon.sm121a.dflash) + DFlash (z-lab/Qwen3.6-35B-A3B-DFlash @31977fbe small-page rev, num_speculative_tokens 11) | NVFP4 (modelopt_mixed — W4A16_NVFP4 experts + FP8 GDN gates) | 65536 | 32 | 407.1 | done | 2026-07-02 07:21 +0800 |
| Qwen3.6-35B-A3B · vLLM-ultimate (AEON) · NVFP4 + DFlash · conc 2 | vLLM (aeon-vllm-ultimate custom container, v0.23.0+aeon.sm121a.dflash) + DFlash (z-lab/Qwen3.6-35B-A3B-DFlash @31977fbe small-page rev, num_speculative_tokens 11) | NVFP4 (modelopt_mixed — W4A16_NVFP4 experts + FP8 GDN gates) | 65536 | 2 | 151.3 | done | 2026-07-02 07:21 +0800 |
| Qwen3.6-35B-A3B · vLLM-ultimate (AEON) · NVFP4 + DFlash · conc 16 | vLLM (aeon-vllm-ultimate custom container, v0.23.0+aeon.sm121a.dflash) + DFlash (z-lab/Qwen3.6-35B-A3B-DFlash @31977fbe small-page rev, num_speculative_tokens 11) | NVFP4 (modelopt_mixed — W4A16_NVFP4 experts + FP8 GDN gates) | 65536 | 16 | 340.2 | done | 2026-07-02 07:21 +0800 |
| Qwen3-Coder-30B-A3B · DDTree vs DFlash · HumanEval (coding workload) · single-stream | DDTree research harness (github.com/liranringel/ddtree, PyTorch + transformers, batch-1) + DDTree tree draft (budgets 64/256) vs single-line DFlash — z-lab/Qwen3-Coder-30B-A3B-DFlash, block_size 16 | BF16 (harness loads the unquantized target via AutoModelForCausalLM) | 4096 | 1 | 49.3 | done | 2026-07-01 21:26 +0800 |
| Qwen3-Coder-30B-A3B · DDTree vs DFlash vs autoregressive · single-stream (batch-1) | DDTree research harness (github.com/liranringel/ddtree, PyTorch + transformers, batch-1) + DDTree tree draft (budgets 64/256) vs single-line DFlash — z-lab/Qwen3-Coder-30B-A3B-DFlash, block_size 16 | BF16 (harness loads the unquantized target via AutoModelForCausalLM) | 4096 | 1 | 20.8 | done | 2026-07-01 21:13 +0800 |
| gpt-oss-120b · vLLM · MXFP4 + EAGLE3 (LMSYS draft) · conc 8 | vLLM + EAGLE3 (lmsys/EAGLE3-gpt-oss-120b-bf16 — the SGLang/SpecForge draft, on vLLM) | MXFP4 | 65536 | 8 | 175.3 | done | 2026-07-01 19:40 +0800 |
| gpt-oss-120b · vLLM · MXFP4 + EAGLE3 (LMSYS draft) · conc 32 | vLLM + EAGLE3 (lmsys/EAGLE3-gpt-oss-120b-bf16 — the SGLang/SpecForge draft, on vLLM) | MXFP4 | 65536 | 32 | 246.7 | done | 2026-07-01 19:21 +0800 |
| gpt-oss-120b · vLLM · MXFP4 + EAGLE3 (LMSYS draft) · conc 1 | vLLM + EAGLE3 (lmsys/EAGLE3-gpt-oss-120b-bf16 — the SGLang/SpecForge draft, on vLLM) | MXFP4 | 65536 | 1 | 25.5 | done | 2026-07-01 17:37 +0800 |
| gpt-oss-20b · vLLM · MXFP4 + EAGLE3 · conc 16 | vLLM + EAGLE3 | MXFP4 | 65536 | 16 | 431.8 | done | 2026-07-01 16:36 +0800 |
| gpt-oss-20b · vLLM · MXFP4 + EAGLE3 · conc 4 | vLLM + EAGLE3 | MXFP4 | 65536 | 4 | 85.4 | done | 2026-07-01 16:15 +0800 |
| gpt-oss-20b · vLLM · MXFP4 + EAGLE3 · conc 2 | vLLM + EAGLE3 | MXFP4 | 65536 | 2 | 59.0 | done | 2026-07-01 16:02 +0800 |
| gpt-oss-20b · vLLM · MXFP4 · conc 16 | vLLM | MXFP4 | 65536 | 16 | 340.7 | done | 2026-07-01 15:35 +0800 |
| gpt-oss-20b · vLLM · MXFP4 · conc 4 | vLLM | MXFP4 | 65536 | 4 | 127.3 | done | 2026-07-01 15:22 +0800 |
| gpt-oss-20b · vLLM · MXFP4 · conc 2 | vLLM | MXFP4 | 65536 | 2 | 83.4 | done | 2026-07-01 15:22 +0800 |
| Qwen3.6-35B-A3B · vLLM · NVFP4 · conc 16 | vLLM | NVFP4 | 65536 | 16 | 332.4 | done | 2026-07-01 14:58 +0800 |
| Qwen3.6-35B-A3B · vLLM · NVFP4 · conc 8 | vLLM | NVFP4 | 65536 | 8 | 241.8 | done | 2026-07-01 14:39 +0800 |
| Qwen3.6-35B-A3B · vLLM · NVFP4 · conc 4 | vLLM | NVFP4 | 65536 | 4 | 173.8 | done | 2026-07-01 14:24 +0800 |
| Qwen3.6-35B-A3B · vLLM · NVFP4 · conc 2 | vLLM | NVFP4 | 65536 | 2 | 113.2 | done | 2026-07-01 14:07 +0800 |
| Qwen3.6-35B-A3B · vLLM · NVFP4 · conc 1 | vLLM | NVFP4 | 65536 | 1 | 74.7 | done | 2026-07-01 13:50 +0800 |
| Qwen3.6-35B-A3B · vLLM · NVFP4 + MTP · conc 16 | vLLM + MTP | NVFP4 | 65536 | 16 | 433.3 | done | 2026-07-01 12:54 +0800 |
| Qwen3.6-35B-A3B · vLLM · NVFP4 + MTP · conc 4 | vLLM + MTP | NVFP4 | 65536 | 4 | 232.4 | done | 2026-07-01 12:37 +0800 |
| Qwen3.6-35B-A3B · vLLM · NVFP4 + MTP · conc 2 | vLLM + MTP | NVFP4 | 65536 | 2 | 161.2 | done | 2026-07-01 12:20 +0800 |
| Qwen3.6-35B-A3B · DDTree (Diffusion Draft Tree) · BLOCKED (hybrid GDN cache) | DDTree research harness (github.com/liranringel/ddtree, PyTorch + transformers, batch-1) + DDTree — tree draft over the z-lab/Qwen3.6-35B-A3B-DFlash block-diffusion drafter | BF16 (harness loads the unquantized target via AutoModelForCausalLM) | 4096 | 1 | — | blocked | 2026-07-01 |
| Qwen3.6-27B (dense) · DDTree (Diffusion Draft Tree) · BLOCKED (hybrid GDN cache) | DDTree research harness (github.com/liranringel/ddtree, PyTorch + transformers, batch-1) + DDTree — tree draft over the z-lab/Qwen3.6-27B-DFlash block-diffusion drafter | BF16 (harness loads the unquantized target via AutoModelForCausalLM) | 4096 | 1 | — | blocked | 2026-07-01 |
| DeepSeek V4-Flash REAP-K180 · ds4 · NVFP4-hybrid · conc 1 (1M ctx) | ds4 (ds4-nvfp4-spark, custom GB10 runtime) + none — ds4 is a serial single-worker; the DeepSeek MTP head is not exercised by this runtime | NVFP4-hybrid (NVFP4 experts + Q2_K + Q8_0 + F16) | 1048576 | 1 | 11.4 | done | 2026-06-28 13:34 +08 |
| MiniMax-M2.7-REAP 172B · vLLM · NVFP4 (W4A4) | vLLM | NVFP4 (W4A4) | 65536 | 32 | 111.9 | done | 2026-06-27 19:22 +0800 |
| MiniMax-M2.7-REAP 172B · vLLM · NVFP4 (W4A4) · conc 1 · 160K | vLLM | NVFP4 (W4A4) | 163840 | 1 | 25.4 | done | 2026-06-27 19:22 +0800 |
| MiniMax-M2.5-REAP 139B · vLLM · NVFP4 (W4A16) · conc 1 · 192K | vLLM | NVFP4 (W4A16) | 196608 | 1 | 27.0 | done | 2026-06-27 15:22 +0800 |
| MiniMax-M2.5-REAP 139B · vLLM · NVFP4 (W4A16) | vLLM | NVFP4 (W4A16) | 65536 | 32 | 120.0 | done | 2026-06-26 22:25 +0800 |
| GLM-4.5-Air-REAP 82B · vLLM · NVFP4 (W4A16) | vLLM | NVFP4 (W4A16) | 65536 | 32 | 158.4 | done | 2026-06-26 16:10 +0800 |
| gpt-oss-120b · SGLang · MXFP4 + EAGLE3 · conc 32 | SGLang + EAGLE3 | MXFP4 | 65536 | 32 | 171.9 | done | 2026-06-23 21:20 +0800 |
| gpt-oss-120b · SGLang · MXFP4 + EAGLE3 · conc 1 | SGLang + EAGLE3 | MXFP4 | 65536 | 1 | 40.6 | done | 2026-06-23 20:08 +0800 |
| Qwen3.6-35B-A3B · vLLM · NVFP4 + MTP · conc 1 | vLLM + MTP | NVFP4 | 65536 | 1 | 93.9 | done | 2026-06-23 13:02 +08 |
| Qwen3.6-35B-A3B · vLLM · NVFP4 + MTP · conc 8 | vLLM + MTP | NVFP4 | 65536 | 8 | 289.1 | done | 2026-06-23 12:49 +08 |
| Qwen3.6-35B-A3B · vLLM · NVFP4 + MTP | vLLM + MTP | NVFP4 | 65536 | 32 | 541.3 | done | 2026-06-23 12:34 +08 |
| Qwen3.6-27B · SGLang · NVFP4 + MTP | SGLang + MTP (NEXTN) | NVFP4 | 65536 | 32 | 196.6 | done | 2026-06-23 12:18 +08 |
| Qwen3.6-35B-A3B · vLLM · NVFP4 | vLLM | NVFP4 | 65536 | 32 | 430.8 | done | 2026-06-23 11:13 +08 |
| gpt-oss-120b · vLLM · MXFP4 + EAGLE3 · conc 8 | vLLM + EAGLE3 | MXFP4 | 65536 | 8 | 50.0 | done | 2026-06-22 23:44 +08 |
| gpt-oss-120b · vLLM · MXFP4 + EAGLE3 · conc 1 | vLLM + EAGLE3 | MXFP4 | 65536 | 1 | 14.7 | done | 2026-06-22 23:29 +08 |
| gpt-oss-20b · vLLM · MXFP4 + EAGLE3 · conc 8 | vLLM + EAGLE3 | MXFP4 | 65536 | 8 | 126.3 | done | 2026-06-22 23:13 +08 |
| gpt-oss-20b · vLLM · MXFP4 + EAGLE3 · conc 1 | vLLM + EAGLE3 | MXFP4 | 65536 | 1 | 38.6 | done | 2026-06-22 23:03 +08 |
| gpt-oss-20b · vLLM · MXFP4 + EAGLE3 | vLLM + EAGLE3 | MXFP4 | 65536 | 32 | 686.5 | done | 2026-06-22 17:52 +08 |
| gpt-oss-120b · vLLM · MXFP4 + EAGLE3 | vLLM + EAGLE3 | MXFP4 | 65536 | 32 | 138.5 | done | 2026-06-22 17:42 +08 |
| gpt-oss-120b · vLLM · MXFP4 | vLLM | MXFP4 | 65536 | 32 | 252.8 | done | 2026-06-22 15:30 +08 |
| gpt-oss-20b · vLLM · MXFP4 | vLLM | MXFP4 | 65536 | 32 | 535.3 | done | 2026-06-22 06:00 +08 |
| Nemotron-3 Nano-4B · llama.cpp · Q4_K_M | llama.cpp | Q4_K_M | 65536 | 32 | 141.3 | done | 2026-06-22 03:10 +08 |
| Nemotron-3 Super 120B · vLLM · NVFP4 | vLLM | NVFP4 | 65536 | 32 | 96.9 | done | 2026-06-22 00:22 +08 |
| Nemotron-3 Nano-Omni 30B-A3B · vLLM · NVFP4 | vLLM | NVFP4 | 65536 | 32 | 388.9 | done | 2026-06-21 23:43 +08 |
| Nemotron-3 Elastic 30B-A3B · vLLM · NVFP4 | vLLM | NVFP4 | 65536 | 32 | 353.0 | done | 2026-06-21 23:23 +08 |
| gpt-oss-120b · SGLang · MXFP4 | SGLang | MXFP4 | 65536 | 32 | 140.3 | done | 2026-06-21 22:56 +08 |
| gpt-oss-20b · SGLang · MXFP4 | SGLang | MXFP4 | 65536 | 32 | 279.5 | done | 2026-06-21 22:08 +08 |
| Qwen3.6-27B (dense) · DDTree (Diffusion Draft Tree) · BLOCKED (hybrid GDN cache) | DDTree research harness (github.com/liranringel/ddtree, PyTorch + transformers, batch-1) + DDTree — tree draft over the z-lab/Qwen3.6-27B-DFlash block-diffusion drafter | BF16 (harness loads the unquantized target via AutoModelForCausalLM) | 4096 | 1 | — | blocked | 2026-07-01 |
| Qwen3.6-35B-A3B · DDTree (Diffusion Draft Tree) · BLOCKED (hybrid GDN cache) | DDTree research harness (github.com/liranringel/ddtree, PyTorch + transformers, batch-1) + DDTree — tree draft over the z-lab/Qwen3.6-35B-A3B-DFlash block-diffusion drafter | BF16 (harness loads the unquantized target via AutoModelForCausalLM) | 4096 | 1 | — | blocked | 2026-07-01 |
| Qwen3.8-27B · SGLang · NVFP4 + DFlash2 · conc-16 | SGLang + DFlash2 | NVFP4 | 262144 | 16 | — | blocked | 2026-08-20 05:25 +0800 |
| Qwen3.8-27B · SGLang · NVFP4 + DFlash2 · conc-32 | SGLang + DFlash2 | NVFP4 | 262144 | 32 | — | blocked | 2026-08-20 05:25 +0800 |
| Qwen3.8-27B · SGLang · NVFP4 + DFlash2 · conc-8 | SGLang + DFlash2 | NVFP4 | 262144 | 8 | — | blocked | 2026-08-20 05:25 +0800 |
| Qwen3.8-27B · SGLang · NVFP4 + DSpark · conc-16 | SGLang + DSpark | NVFP4 | 262144 | 16 | — | blocked | 2026-08-20 05:25 +0800 |
| Qwen3.8-27B · SGLang · NVFP4 + DSpark · conc-32 | SGLang + DSpark | NVFP4 | 262144 | 32 | — | blocked | 2026-08-20 05:25 +0800 |
| gpt-oss-20b · llama.cpp · MXFP4 | llama.cpp | MXFP4 | 131072 | 32 | — | blocked | 2026-06-21 |