NVFP4 (unsloth; compressed-tensors) + fp8 KV + triton_attn (fp8-KV auf GB10 nur mit triton, FA kann kein fp8-KV auf sm_121)vLLMWeights Repo Updated Aug 15, 2026

Details

2× DGX Spark

Qwen3.8-Flash-Next NVFP4 with vLLM on 2× DGX Spark (Weschera)

by Weschera

33tok/sEveryday

1 request, prose prompt, with MTP Source

Intelligenceno independent value

Custom kernel required
License unclear

NVFP4 (ModelOpt MIXED_PRECISION: NVFP4 Experts, FP8 dense/PLE/MTP); BF16 KV (Pflicht — QSA lehnt fp8-KV ab)vLLMWeights Repo Updated Sep 5, 2026

Details

1× Mac

DeepSeek-V4-Flash 284B IQ2_XXS with DwarfStar on 1× Mac (vision)

by Weschera

28.5tok/sEveryday

1 request, prose prompt, context 32k, with DSpark (per recipe) Source

Intelligenceno independent value

≤ 3 bit: quantization may cost quality

IQ2_XXS imatrix q2 (antirez), w2Q2K/AProjQ8/SExpQ8/OutQ8; GGUF 86,720,111,776 bytesds4 (DwarfStar)Weights Repo Updated Sep 2, 2026

Details

2× DGX Spark

Ling-3.0-flash-VL FP8 with SGLang on 2× DGX Spark

by Weschera

26tok/sEveryday

1 request, prose prompt, no speculative decoding Source

Intelligenceno independent value

FP8 (official inclusionAI FP8-Checkpoint; ~118 GiB)SGLangWeights Repo Updated Sep 9, 2026

Details

1× DGX Spark

Qwen3.8-Flash-Next 125B UD-Q4_K_XL with llama.cpp on 1× DGX Spark

by Weschera

24.2tok/sEveryday

1 request, prose prompt, with Draft Source

Intelligenceno independent value

License unclear

Unsloth UD-Q4_K_XL GGUF (103,7 GiB, 4 Shards; PLE ≥4-bit) + MTP-Draft shared-Q8_0 (2,6 GiB)llama.cppWeights Repo Updated Sep 2, 2026

Details

2× DGX Spark

GLM-5.3-Flash NVFP4 with vLLM on 2× DGX Spark (128k context, with MTP)

by Weschera

23tok/sEveryday

1 request, prose prompt, with MTP4 Source

Intelligenceno independent value

Experimental
Custom kernel required

NVFP4 (vcruz305-Quant, BF16-MTP-Head, 49 Shards; quantization_config.ignore für Layer 45), KV fp8_e4m3vLLMWeights Repo Updated Aug 28, 2026

Details

1× DGX Spark

Spark-X2.5-4B BF16 with SGLang on 1× DGX Spark

by Weschera

21.8tok/sEveryday

1 request, prose prompt, no speculative decoding Source

Intelligenceno independent value

BF16 (unquantisiert)SGLangWeights Repo Updated Sep 7, 2026

Details

2× DGX Spark

GLM-5.3-Flash NVFP4 with vLLM on 2× DGX Spark (128k context, with DFlash2)

by Weschera

20.5tok/sEveryday

1 request, prose prompt, no speculative decoding Source

Intelligenceno independent value

Non-commercial

NVFP4 (Marlin W4A16); FP8 E4M3 KV (6 GiB-Pin)vLLMWeights Repo Updated Sep 10, 2026

Details

1× DGX Spark

Laguna-S-2.1 118B NVFP4 with vLLM on 1× DGX Spark (128k context)

by Weschera

19.3tok/sEveryday

1 request, prose prompt, context 1k, no speculative decoding Source

Intelligenceno independent value

License unclear

NVFP4 (Blackwell-native, auto-detected aus quantization_config)vLLMWeights Repo Updated Jul 21, 2026

Details

1× DGX Spark

Qwen3.8-27B NVFP4 with Atlas on 1× DGX Spark

by Weschera

19tok/sEveryday

1 request, prose prompt, with MTP (per recipe) Source

Intelligenceno independent value

Experimental
Custom kernel required

NVFP4 (unsloth); MTP-Head bf16 (--mtp-quantization bf16); fp8 KV (Engine-Default)AtlasWeights Repo Updated Aug 15, 2026

Details

1× DGX Spark

GLM-5.3-Flash 320B UD-Q2_K_XL with llama.cpp on 1× DGX Spark

by Weschera

17.9tok/sEveryday

1 request, prose prompt, context 32k, no speculative decoding Source

Intelligenceno independent value

≤ 3 bit: quantization may cost quality

Unsloth UD-Q2_K_XL (109 GB, ~2,4-bit Experts dynamisch; Attention/Router höher)llama.cppWeights Repo Updated Aug 27, 2026

Details

1× MacBonsai-2-27B (Qwen3.8-27B-Basis)

Bonsai-2-27B Q8_0 with llama.cpp on 1× Mac

by Weschera

10.6tok/sEveryday

1 request, prose prompt, no speculative decoding Source

Intelligenceno independent value

Custom kernel required

Ternary PQ2_0 (empfohlen) / PTQ1_0 (5,9 GB); mmproj-Q8_0 für Visionllama.cpp (prism fork)Weights Repo Updated Sep 18, 2026

Details

2× DGX Spark

DeepSeek-V4-Flash FP8 with vLLM on 2× DGX Spark (1024k context)

by Weschera

83.8tok/sMixed

1 request, mixed prompt set, with Draft Source

Intelligenceno independent value

Official BF16/FP8-Checkpoint mit nvfp4_ds_mla-KV (block 256); MoE FlashInfer B12XvLLMWeights Repo Updated Jul 31, 2026

Details

2× DGX Spark

Qwen3.8-Flash-Next NVFP4 with SGLang on 2× DGX Spark

by Weschera

41.7tok/sMixed

1 request, mixed prompt set, with NEXTN (MTP) (per recipe) Source

Intelligenceno independent value

Custom kernel required
License unclear

NVFP4 (modelopt_fp4, RadixArk-Checkpoint); BF16 SSM; NEXTN-Draft in-checkpointSGLangWeights Repo Updated Aug 27, 2026

Details

1× DGX Spark

Qwen3.8-Flash-Next 80B UD-IQ4_XS with llama.cpp on 1× DGX Spark

by Weschera

26.9tok/sMixed

1 request, mixed prompt set, no speculative decoding Source

Intelligenceno independent value

License unclear

Unsloth UD-IQ4_XS GGUF (94 GB)llama.cppWeights Repo Updated Aug 27, 2026

Details

1× DGX Spark

Qwen3.8-27B NVFP4 with SGLang on 1× DGX Spark (Weschera)

by Weschera

25.2tok/sMixed

1 request, mixed prompt set, with DFlash2 (per recipe) Source

Intelligenceno independent value

Custom kernel required

NVFP4 (RadixArk) + BF16-Draft; quantized lm_head-Adapter (TP=1, 256-token selector chunks, FP32 top-k — r0b0tlab-Ansatz)SGLangWeights Repo Updated Aug 21, 2026

Details

1× DGX Spark

GLM-5.3-Flash 320B UD-IQ3_XXS with llama.cpp on 1× DGX Spark

by Weschera

20.8tok/sMixed

1 request, mixed prompt set, with MTP Source

Intelligenceno independent value

≤ 3 bit: quantization may cost quality

Unsloth Dynamic UD-IQ3_XXS GGUF (120,4 GB, 4 Shards); KV q8_0llama.cppWeights Repo Updated Sep 4, 2026

Details

1× DGX Spark

Qwen3.6-35B NVFP4 with SGLang on 1× DGX Spark

by Weschera

92.2tok/sPeak

1 request, best value reported by source, context 64k Source

Intelligenceno independent value

NVFP4 (qwen36-35b-nvfp4 Checkpoint)SGLangWeights Repo Updated Jul 12, 2026

Details

1× Mac

Qwen3.8-Flash-Next 125B with MLX on 1× Mac

by Weschera

83.1tok/sPeak

1 request, code prompt Source

Intelligenceno independent value

License unclear

oQ4e gemischt (4-bit gesamt, group 64, MLX safetensors)oMLX 0.6.4 (custom kernels)Weights Repo Updated Aug 30, 2026

Details

2× DGX Spark

DeepSeek-V4-Flash NVFP4 with vLLM on 2× DGX Spark

by Weschera

65.4tok/sPeak

1 request, best value reported by source Source

Intelligenceno independent value

Experimental
Custom kernel required

NVFP4-KV-Pfad (Stage C: 584-byte padded envelope via nvfp4_ds_mla; NICHT der 416-byte true-layout Fix — dokumentiert gescheitert >411 prompt tokens)vLLMWeights Repo Updated Jul 2, 2026

Details

4× DGX Spark

GLM-5.2 753B INT4/INT8 with vLLM on 4× DGX Spark

by Weschera

32.5tok/sPeak

1 request, best value reported by source, context 200k Source

Intelligenceno independent value

Experimental
Custom kernel required

QuantTrio Int4-Int8Mix (W4A16 MoE-Experts / W8A16 dense+MTP, compressed-tensors, data-free; Layer 0 + Attention-Indexer BF16); KV nvfp4_ds_mla (400 B/Token)vLLMWeights Repo Updated Aug 20, 2026

Details

4× DGX Spark