Local AI · Recipe · 1× RTX 3090

Qwen3.8-27B EXL3 4.00 bpw with buun-llama-cpp on 1× RTX 3090

Qwen3.8-27B EXL3 4.00 bpw with buun-llama-cpp on 1× RTX 3090: 27.8 tok/s according to github.com (dataset as of Nov 3, 2026).

by r0b0tlab

27.8tok/sMixed

1 request, mixed prompt set, context 150000, with DFlash2 (per recipe)

Source: github.com

Intelligence (original model) · Artificial Analysis34with thinking no thinking 20

Instructions in the README Repo

Engine: buun-llama-cpp
Quantization: EXL3 4.00 bpw (Target + DFlash2-Draft), VBR-KV turbo8/turbo3
Model family: Qwen3.8-27B
Context: 262,144
Parameters: 27B dense
Creator: r0b0tlab
GitHub stars: 4
Repo updated: Sep 20, 2026

Measurements

Every sourced value of this recipe, each with its condition and source. Bars relative to the largest value in the group.

Decode · 1 request

Mixed (decode)27.8 tok/scontext 146.5K · Decode nach 150000-Token-Prefill Source
Peak (decode)149 tok/sGSM8K throughput n=40: mean E2E tok/s pro Request; mean AL 5.917 (Concurrency: one active slot laut Rezept) Source

Prefill by context

146.5K525 tok/s150000-token prefill + 200 decode Source

Time to first token

Shorter is better.

Prompt not stated175 msMixed short requests median TTFT 0.175 s Source

What you need

Hardware: 1 × NVIDIA RTX 3090
Weights: Main weightsr0b0tlab/Qwen3.8-27B-EXL3-4.00bpw Apache-2.0
Draft modelr0b0tlab/Qwen3.8-27B-DFlash2-EXL3-4.00bpw Apache-2.0
Engine: buun-llama-cpp
Context: 262,144 tokens

Notes

What matters before you rebuild it.

Custom kernel required

Sources

Related recipes

1× Mac

Qwen3.8-27B with MLX on 1× Mac

by Weschera

53.3tok/sEveryday

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

Intelligenceno independent value

License unclear

oQ4e (4-bit affine g64, 166 sens. Tensoren 5-bit; Stock-Qwen3.8-27B-Konversion)oMLX 0.6.3rc2Weights Repo Updated Aug 21, 2026

Details

1× DGX Spark

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

by Weschera

45tok/sEveryday

1 request, prose prompt, with MTP Source

Intelligenceno independent value

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

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

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

by 0xBakeer

49.9tok/sMixed

1 request, benchmark prompt set, with DFlash2 drafter + StairCut (per recipe) Source

Intelligence34with thinking no thinking 20

Custom kernel required

NVFP4 (GPTQ) alle Projektionen + FP8-Head (über Qwen FP8-Checkpoint)TandemLLM v0.3.0 (eigene PyTorch/Triton/CUDA-Engine, StairCut)Weights Repo Updated Oct 30, 2026

Details

1× DGX Spark

Qwen3.8-27B NVFP4 with vLLM on 1× DGX Spark (r0b0tlab)

by r0b0tlab

27.8tok/sMixed

1 request, mixed prompt set, with MTP Source

Intelligence34with thinking no thinking 20

License unclear

NVFP4 (+ BF16-MTP-Kopf)vLLM v0.27.2rc0-sm121 (7f7a32c)Weights Repo Updated Aug 20, 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

← Back to overview

Local AI in your company?

In a workshop we work out which models and which hardware fit your tasks, and build the first agent on your infrastructure.

See the AI workshop Back to overview