| **Qwen 3.5 0.8B** | <a href="https://huggingface.co/Qwen/Qwen3.5-0.8B/blob/main/model.safetensors-00001-of-00001.safetensors"><img src="https://img.shields.io/badge/target-BF16-f5c842?style=flat-square&labelColor=090909" alt="target: BF16"></a> | <a href="https://www.lucebox.com/blog/megakernel"><img src="https://img.shields.io/badge/Megakernel-1c1c1f?style=flat-square&labelColor=1c1c1f" alt="Megakernel"></a> | Prefill + decode | **1.9×** prefill; **1.55×** decode |
| **Qwen 3.8 27B**<br><sub>on R9700</sub> | <a href="https://huggingface.co/unsloth/Qwen3.8-27B-GGUF/blob/main/Qwen3.8-27B-UD-IQ4_XS.gguf"><img src="https://img.shields.io/badge/target-UD--IQ4__XS-f5c842?style=flat-square&labelColor=090909" alt="target: UD-IQ4_XS"></a><br><a href="https://huggingface.co/incoai/Qwen3.8-27B-DFlash2/blob/main/model.safetensors"><img src="https://img.shields.io/badge/drafter-DFlash2%20source%20%E2%86%92%20Q8__0-e8e8ed?style=flat-square&labelColor=090909" alt="drafter: DFlash2 source → Q8_0"></a> | <a href="https://www.lucebox.com/blog/qwen38-r9700"><img src="https://img.shields.io/badge/DFlash2-1c1c1f?style=flat-square&labelColor=1c1c1f" alt="DFlash2"></a> | Decode | **6.4×** vs Lucebox AR; **3.8×** vs llama.cpp with the same drafter |
| **Laguna XS 2.1 33B** | <a href="https://huggingface.co/poolside/Laguna-XS-2.1-GGUF/blob/main/Laguna-XS-2.1-Q4_K_M.gguf"><img src="https://img.shields.io/badge/target-Q4__K__M-f5c842?style=flat-square&labelColor=090909" alt="target: Q4_K_M"></a><br><a href="https://huggingface.co/Qwen/Qwen3-0.6B-GGUF/blob/main/Qwen3-0.6B-Q8_0.gguf"><img src="https://img.shields.io/badge/scorer-Qwen3%200.6B%20Q8__0-e8e8ed?style=flat-square&labelColor=090909" alt="scorer: Qwen3 0.6B Q8_0"></a> | <a href="https://www.lucebox.com/blog/pflash"><img src="https://img.shields.io/badge/PFlash-1c1c1f?style=flat-square&labelColor=1c1c1f" alt="PFlash"></a> <a href="https://www.lucebox.com/blog/kvflash"><img src="https://img.shields.io/badge/KVFlash-1c1c1f?style=flat-square&labelColor=1c1c1f" alt="KVFlash"></a> | Prefill | **6.1×**, **411 s to 67.3 s** at 256K |
| **Laguna XS 2.1 33B** | <a href="https://huggingface.co/poolside/Laguna-XS-2.1-GGUF/blob/main/Laguna-XS-2.1-Q4_K_M.gguf"><img src="https://img.shields.io/badge/target-Q4__K__M-f5c842?style=flat-square&labelColor=090909" alt="target: Q4_K_M"></a><br><a href="https://huggingface.co/Lucebox/Laguna-XS-2.1-DFlash-GGUF/blob/main/laguna-xs21-dflash-q4.gguf"><img src="https://img.shields.io/badge/drafter-DFlash%20Q4-e8e8ed?style=flat-square&labelColor=090909" alt="drafter: DFlash Q4"></a> | <a href="https://www.lucebox.com/blog/dflash27b"><img src="https://img.shields.io/badge/DFlash-1c1c1f?style=flat-square&labelColor=1c1c1f" alt="DFlash"></a> | Decode | **1.7×** at 256K |
| **Gemma 4 26B-A4B** | <a href="https://huggingface.co/bartowski/google_gemma-4-26B-A4B-it-GGUF/blob/main/google_gemma-4-26B-A4B-it-Q4_K_M.gguf"><img src="https://img.shields.io/badge/target-Q4__K__M-f5c842?style=flat-square&labelColor=090909" alt="target: Q4_K_M"></a><br><a href="https://huggingface.co/Lucebox/gemma-4-26B-A4B-it-DFlash-GGUF/blob/main/gemma-4-26B-A4B-it-DFlash-q8_0.gguf"><img src="https://img.shields.io/badge/drafter-DFlash%20Q8__0-e8e8ed?style=flat-square&labelColor=090909" alt="drafter: DFlash Q8_0"></a> | <a href="https://www.lucebox.com/blog/dflash27b"><img src="https://img.shields.io/badge/DFlash-1c1c1f?style=flat-square&labelColor=1c1c1f" alt="DFlash"></a> | Decode | **1.31×** |
| **Gemma 4 31B IT** | <a href="https://huggingface.co/bartowski/google_gemma-4-31B-it-GGUF/blob/main/google_gemma-4-31B-it-Q4_K_M.gguf"><img src="https://img.shields.io/badge/target-Q4__K__M-f5c842?style=flat-square&labelColor=090909" alt="target: Q4_K_M"></a><br><a href="https://huggingface.co/Lucebox/gemma-4-31B-it-DFlash-GGUF/blob/main/gemma-4-31B-it-DFlash-q8_0.gguf"><img src="https://img.shields.io/badge/drafter-DFlash%20Q8__0-e8e8ed?style=flat-square&labelColor=090909" alt="drafter: DFlash Q8_0"></a> | <a href="https://www.lucebox.com/blog/dflash27b"><img src="https://img.shields.io/badge/DFlash-1c1c1f?style=flat-square&labelColor=1c1c1f" alt="DFlash"></a> | Decode | **3.2×** |
| **DeepSeek V4 Flash**<br><sub>on Strix Halo</sub> | <a href="https://huggingface.co/Lucebox/DeepSeek-V4-Flash-0731-ROCmFP3/blob/main/DeepSeek-V4-Flash-0731-ROCMFPX-MIX-STRIX.gguf"><img src="https://img.shields.io/badge/target-ROCmFPX%20MIX%20Strix-f5c842?style=flat-square&labelColor=090909" alt="target: ROCmFPX MIX Strix"></a><br><a href="https://huggingface.co/Lucebox/DeepSeek-V4-Flash-0731-DSpark-GGUF/blob/main/DeepSeek-V4-Flash-0731-DSpark-draft-Q4RMFP4-denseF16.gguf"><img src="https://img.shields.io/badge/drafter-DSpark%20Q4RMFP4-e8e8ed?style=flat-square&labelColor=090909" alt="drafter: DSpark Q4RMFP4"></a> | <a href="https://www.lucebox.com/blog/deepseek-v4-flash-0731"><img src="https://img.shields.io/badge/DSpark-1c1c1f?style=flat-square&labelColor=1c1c1f" alt="DSpark"></a> | Decode | **42 tok/s** at 8K and **39 tok/s** on code and math with the plain launch ([PR #729](https://github.com/Luce-Org/lucebox/pull/729)) |
**DeepSeek V4.1 Flash**<br><sub>R9700 + Strix Halo + SSD, --profile ds41-lucebox</sub> | <a href="https://huggingface.co/Lucebox/DeepSeek-V4.1-Flash-1017-ROCMFP-GGUF/blob/main/DeepSeek-V4.1-Flash-1017-ROCMFP2S-IM1017.gguf"><img src="https://img.shields.io/badge/target-ROCMFP2S-f5c842?style=flat-square&labelColor=090909" alt="target: ROCMFP2S"></a><br><a href="https://huggingface.co/Lucebox/DeepSeek-V4.1-Flash-1017-ROCMFP-GGUF/blob/main/DeepSeek-V4.1-Flash-1017-DSpark-draft-MXFP4-Q8.gguf"><img src="https://img.shields.io/badge/drafter-DSpark%20MXFP4--Q8-e8e8ed?style=flat-square&labelColor=090909" alt="drafter: DSpark MXFP4-Q8"></a> | <a href="https://www.lucebox.com/blog/deepseek-v4-flash-0731"><img src="https://img.shields.io/badge/DSpark-1c1c1f?style=flat-square&labelColor=1c1c1f" alt="DSpark"></a> | Decode | **25.4-25.8 tok/s** on code, **16.8-17.5 tok/s** after a 4K prompt, at the profile's 128K context ([guide](server/docs/DS41.md)) |
**DeepSeek V4.1 Flash**<br><sub>R9700 + Ryzen AI Max+ PRO 495 192 GB, --profile ds41-gorgon</sub> | <a href="https://huggingface.co/Lucebox/DeepSeek-V4.1-Flash-1017-ROCMFP-GGUF/blob/main/DeepSeek-V4.1-Flash-1017-ROCMFP2S-IM1017.gguf"><img src="https://img.shields.io/badge/target-ROCMFP2S-f5c842?style=flat-square&labelColor=090909" alt="target: ROCMFP2S"></a><br><a href="https://huggingface.co/Lucebox/DeepSeek-V4.1-Flash-1017-ROCMFP-GGUF/blob/main/DeepSeek-V4.1-Flash-1017-DSpark-draft-Q2K-Q4K.gguf"><img src="https://img.shields.io/badge/drafter-DSpark%20Q2K--Q4K-e8e8ed?style=flat-square&labelColor=090909" alt="drafter: DSpark Q2K-Q4K"></a> | <a href="https://www.lucebox.com/blog/deepseek-v4-flash-0731"><img src="https://img.shields.io/badge/DSpark-1c1c1f?style=flat-square&labelColor=1c1c1f" alt="DSpark"></a> | Prefill + decode | **985-1,019 tok/s** prefill at 14K-25K, **48.7 tok/s** decode on code, every expert resident ([guide](server/docs/DS41.md#running-on-an-r9700-with-a-192-gb-ryzen-ai-max)) |
| **Ling 3.0 Flash 124B-A5.1B**<br><sub>on DGX Spark</sub> | <a href="https://huggingface.co/bloomer010/Ling-3.0-flash-GGUF"><img src="https://img.shields.io/badge/target-Q4__K__M-f5c842?style=flat-square&labelColor=090909" alt="target: Q4_K_M"></a> | Autoregressive | Decode | **34.6 tok/s** median AR |
| **Qwen 3.8 Flash-Next**<br><sub>on Strix Halo</sub> | <a href="https://huggingface.co/Lucebox/Qwen3.8-Flash-Next-GGUF/blob/main/UD-Q4_K_XL/Qwen3.8-Flash-Next-UD-Q4_K_XL-00001-of-00004.gguf"><img src="https://img.shields.io/badge/target-UD--Q4__K__XL-f5c842?style=flat-square&labelColor=090909" alt="target: UD-Q4_K_XL"></a><br><a href="https://huggingface.co/Lucebox/Qwen3.8-Flash-Next-GGUF/blob/main/MTP/mtp-Qwen3.8-Flash-Next-shared-Q8_0.gguf"><img src="https://img.shields.io/badge/drafter-MTP%20Q8__0-e8e8ed?style=flat-square&labelColor=090909" alt="drafter: MTP Q8_0"></a> | <a href="server/docs/QWEN4EXP.md"><img src="https://img.shields.io/badge/MTP-1c1c1f?style=flat-square&labelColor=1c1c1f" alt="MTP"></a> | Decode | **40 tok/s** on a short reply, **1.8×** autoregressive decode ([guide](server/docs/QWEN4EXP.md)) |
**Qwen 3.8 Flash-Next**<br><sub>R9700 + Strix Halo, --profile qwen-next-r9700-strix</sub> | <a href="https://huggingface.co/Lucebox/Qwen3.8-Flash-Next-GGUF/blob/main/UD-Q4_K_XL/Qwen3.8-Flash-Next-UD-Q4_K_XL-00001-of-00004.gguf"><img src="https://img.shields.io/badge/target-UD--Q4__K__XL-f5c842?style=flat-square&labelColor=090909" alt="target: UD-Q4_K_XL"></a><br><a href="https://huggingface.co/Lucebox/Qwen3.8-Flash-Next-GGUF/blob/main/MTP/mtp-Qwen3.8-Flash-Next-shared-Q8_0.gguf"><img src="https://img.shields.io/badge/drafter-MTP%20Q8__0-e8e8ed?style=flat-square&labelColor=090909" alt="drafter: MTP Q8_0"></a> | <a href="server/docs/QWEN4EXP.md"><img src="https://img.shields.io/badge/MTP-1c1c1f?style=flat-square&labelColor=1c1c1f" alt="MTP"></a> <a href="server/docs/QWEN4EXP.md#running-on-an-r9700--strix-halo"><img src="https://img.shields.io/badge/Expert%20split-1c1c1f?style=flat-square&labelColor=1c1c1f" alt="Expert split"></a> | Prefill + decode | **75 tok/s** decode, **1.86×** the Strix Halo alone; **1,820 tok/s** prefill at 16K ([guide](server/docs/QWEN4EXP.md#running-on-an-r9700--strix-halo)) |
| **Qwen 3.8 27B Vision**<br><sub>on R9700</sub> | <a href="https://huggingface.co/Lucebox/Qwen3.8-27B-IQ4_XS-fast-GGUF/blob/main/Qwen3.8-27B-IQ4_XS-pure.gguf"><img src="https://img.shields.io/badge/target-IQ4__XS-f5c842?style=flat-square&labelColor=090909" alt="target: IQ4_XS"></a><br><a href="https://huggingface.co/Lucebox/Qwen3.8-27B-IQ4_XS-fast-GGUF/blob/main/Qwen3.8-27B-mmproj-Q8_0.gguf"><img src="https://img.shields.io/badge/vision-mmproj%20Q8__0-8ab4f8?style=flat-square&labelColor=090909" alt="vision: mmproj Q8_0"></a><br><a href="https://huggingface.co/Lucebox/Qwen3.8-27B-DFlash2-GGUF/blob/main/Qwen3.8-27B-DFlash2-Q8_0.gguf"><img src="https://img.shields.io/badge/drafter-DFlash2%20Q8__0-e8e8ed?style=flat-square&labelColor=090909" alt="drafter: DFlash2 Q8_0"></a> | <a href="https://www.lucebox.com/blog/qwen38-r9700"><img src="https://img.shields.io/badge/DFlash2-1c1c1f?style=flat-square&labelColor=1c1c1f" alt="DFlash2"></a> <a href="https://www.lucebox.com/blog/vision-llm-inference"><img src="https://img.shields.io/badge/Image%20input-1c1c1f?style=flat-square&labelColor=1c1c1f" alt="Image input"></a> | Image questions | **2.4×** vs llama.cpp on DGX Spark at 8 users (10.5 s vs 25.4 s); **2.1×** for one user |
| **DeepSeek V4 Flash Vision**<br><sub>encoder on R9700, 4 users at once</sub> | <a href="https://huggingface.co/Lucebox/DeepSeek-V4-Flash-0731-ROCmFP3/blob/main/DeepSeek-V4-Flash-Vision-Exp-ROCMFPX-MIX-STRIX.gguf"><img src="https://img.shields.io/badge/target-ROCmFPX%20MIX%20Strix-f5c842?style=flat-square&labelColor=090909" alt="target: ROCmFPX MIX Strix"></a><br><a href="https://huggingface.co/Lucebox/DeepSeek-V4-Flash-0731-ROCmFP3/blob/main/DeepSeek-V4-Flash-Vision-Exp-mmproj-BF16.gguf"><img src="https://img.shields.io/badge/vision-mmproj%20BF16-8ab4f8?style=flat-square&labelColor=090909" alt="vision: mmproj BF16"></a> | <a href="https://www.lucebox.com/blog/vision-llm-inference"><img src="https://img.shields.io/badge/Image%20input-1c1c1f?style=flat-square&labelColor=1c1c1f" alt="Image input"></a> | Image questions | **1.5×** sooner first token with 16 images when the R9700 encodes |
| **MOSS-Image SFT 5B**<br><sub>on R9700, text to image</sub> | <a href="https://huggingface.co/Comfy-Org/MOSS-Image/blob/main/diffusion_models/moss_image_sft_bf16.safetensors"><img src="https://img.shields.io/badge/target-BF16%20source%20%E2%86%92%20int8%20convrot-f5c842?style=flat-square&labelColor=090909" alt="target: BF16 source → int8 convrot"></a><br><a href="https://huggingface.co/Lucebox/Qwen3.5-9B-GGUF/blob/main/Qwen3.5-9B-Q8_0.gguf"><img src="https://img.shields.io/badge/text%20encoder-Qwen3.5%209B%20Q8__0-e8e8ed?style=flat-square&labelColor=090909" alt="text encoder: Qwen3.5 9B Q8_0"></a><br><a href="https://huggingface.co/Comfy-Org/MOSS-Image/blob/main/vae/flux2-vae.safetensors"><img src="https://img.shields.io/badge/VAE-FLUX.2-8ab4f8?style=flat-square&labelColor=090909" alt="VAE: FLUX.2"></a> | <a href="optimizations/image_gen/README.md"><img src="https://img.shields.io/badge/Native%20engine%2C%20int8%20kernels-1c1c1f?style=flat-square&labelColor=1c1c1f" alt="Native engine, int8 kernels"></a> | Image generation | **3.1 s** per 1024×1024 image at 8 steps in luce_server; **1.5-1.8×** a DGX Spark on ComfyUI ([results](optimizations/image_gen/RESULTS.md)) |