Will Qwen3.6, Gemma 4 or gpt-oss run on your PC?
Real measured tokens/sec from 105 home AI rigs, each linked to its source. No guesses, no marketing numbers.
Check my GPU (VRAM calculator) Best GPU in 2026 → · Best value per € → · Mac vs PC → · Used 3090 guide → · AMD Radeon → · DGX Spark vs Mac Studio vs Strix Halo → · 3090 vs 5090 → · Best GPU under €1,000 → · All comparisons →
New: What GPU do you need for Qwen3.6, Gemma 4 and gpt-oss? (measured) · Can X run on Y? · All builds A–Z · By GPU · By VRAM size · Intel Arc Pro B70 · Free dataset (CC BY 4.0)
Popular 2026 models: Qwen3.8 27B · Qwen3.6 35B-A3B · Qwen3.6 27B · Muse Spark Open 816B-A42B · Gemma 4 31B · Gemma 4 26B-A4B · Gemma 4 12B · gpt-oss 120B · gpt-oss 20B · Qwen3.5 122B · Qwen3.5 9B · Mistral Medium 3.5 128B · Muse Glimmer 30B · Nemotron 3 Nano Omni 30B · GLM 4.7 Flash · LFM2.5 24B-A2B · Nemotron 3.5 Bolt 74B-A9B · Nemotron 3.5 Super VL 120B-A12B · Qwen3.8 Flash-Next · Flint 40B-A2.5B · Nemotron 3.5 Lightning 30B-A3B · Nemotron 3 Super 120B-A12B · MiniMax M2.7 · DeepSeek V4 Flash · DeepSeek V4.1 Flash · GLM-5.3 Flash · Ornith 1.5 122B · GLM-5.2 · Ornith 1.5 35B-A3B · Qwen3 235B-A22B · DeepSeek R1 671B
Max budget (€): Target model: + Submit your build
| Build | GPU(s) | VRAM GB | Cost | Measured tok/s |
|---|---|---|---|---|
| Mac Studio M3 Ultra, 512 GB ⚠ unchecked | Apple M3 Ultra | 512 | DeepSeek R1 671B: 18.43 | |
| Mac Studio M3 Ultra 512GB (Geerling test) | Apple M3 Ultra (unified) | 512 | DeepSeek R1 671B: 19.89 gpt-oss 20B: 115.29 gpt-oss 20B: 115.52 Llama 3.1 70B: 14.08 Llama 3.1 405B: 1.96 DeepSeek R1 Distill 14B: 51.85 Llama 3.2 3B: 154.6 | |
| Mac Studio M3 Ultra 512GB: Muse Spark Open 816B (Hugging Face test) | Apple M3 Ultra (unified) | 512 | Muse Spark Open 816B-A42B: 18 Muse Spark Open 816B-A42B: 21 | |
| Mac Studio M5 Ultra 512GB: Qwen3.8 27B and Mistral Medium 3.5 (Apple MLX benchmark) | Apple M5 Ultra (unified) | 512 | Qwen3.8 27B: 19.69 Mistral Medium 3.5 128B: 7.7 | |
| Mac Studio M3 Ultra 512GB: Qwen3.8 27B and Mistral Medium 3.5 (Apple MLX benchmark) | Apple M3 Ultra (unified) | 512 | Qwen3.8 27B: 12.46 Mistral Medium 3.5 128B: 4.95 | |
| 4x DGX Spark cluster running GLM-5.2 (unpruned) | 4x Nvidia GB10 (unified, DGX Spark) | 512 | GLM-5.2: 32.5 ⚡ | |
| 4x DGX Spark ring running GLM-5.3 Flash NVFP4 | 4x Nvidia GB10 (unified, DGX Spark) | 512 | GLM-5.3 Flash: 61.7 GLM-5.3 Flash: 105.3 | |
| 2x DGX Spark running GLM-5.3 Flash NVFP4 | 2x Nvidia GB10 (unified, DGX Spark) | 256 | GLM-5.3 Flash: 33.0 GLM-5.3 Flash: 56.5 | |
| Mac Studio M2 Ultra 76-core GPU, 192 GB | Apple M2 Ultra 76-core GPU | 192 | Llama 2 7B: 125.21 gpt-oss 20B: 116.08 gpt-oss 120B: 79.68 | |
| Minisforum MS-S1 MAX (Ryzen AI Max+ PRO 495, 192 GB): gpt-oss 120B at 55.8 tok/s | AMD Ryzen AI Max+ PRO 495 (Radeon 8065S iGPU, 40 CUs) | 192 | $7,399 (≈€6,602) | gpt-oss 120B: 55.8 gpt-oss 120B: 45.1 MiniMax M2.7: 29.5 MiniMax M2.7: 17.3 DeepSeek V4 Flash: 17.4 DeepSeek V4 Flash: 12.2 Qwen3.8 27B: 8 |
| 8x RTX 3090 serving DeepSeek V4 Flash (LimeChain vLLM kit) | 8x RTX 3090 24 GB | 192 | DeepSeek V4 Flash: 95.52 | |
| Strix Halo + Radeon AI Pro R9700 over OCuLink: Qwen3.8 Flash-Next at ~102 tok/s (custom engine) | AMD Radeon AI Pro R9700 32 GB (OCuLink x4) + Ryzen AI Max+ 395 Radeon 8060S iGPU | 160 | Qwen3.8 Flash-Next: 102 ⚡ Qwen3.8 Flash-Next: 82 Qwen3.8 Flash-Next: 71 Qwen3.8 Flash-Next: 29-34 | |
| EPYC 7F52 jank rig: 4x3090 + 4090 + 4080 (136 GB VRAM) ⚠ unchecked | 4x RTX 3090, RTX 4090, RTX 4080 | 136 | Llama 3.1 70B: 9 | |
| Mac Studio M1 Ultra (64-core GPU) 128GB | Apple M1 Ultra 64-core GPU (unified) | 128 | gpt-oss 20B: 100.43 DeepSeek R1 Distill 14B: 41.51 Llama 3.1 70B: 9.84 Llama 3.2 3B: 127.6 | |
| HP OmniBook Ultra 16 (RTX Spark N1X, 128GB) laptop: local AI benchmarks | Nvidia RTX Spark N1X, 6,144-core Blackwell GPU (unified) | 128 | $6,230 (≈€5,559) | Qwen3.6 35B-A3B: 89.5 gpt-oss 20B: 58.0 Llama 3.1 8B: 56.5 |
| DGX Spark (GB10, 128GB): Nemotron 3.5 Super VL 120B and Bolt 74B, launch-week runs | Nvidia GB10 (unified) | 128 | Nemotron 3.5 Super VL 120B-A12B: 19.06 Nemotron 3.5 Super VL 120B-A12B: 38.07 ⚡ Nemotron 3.5 Bolt 74B-A9B: 39.47 Nemotron 3.5 Bolt 74B-A9B: 61.61 ⚡ | |
| Strix Halo (Corsair AI Workstation 300, 128 GB): ROCm vs Vulkan on Qwen3.8 27B and Qwen3.6 35B-A3B | AMD Ryzen AI Max+ 395 (Radeon 8060S iGPU) | 128 | Qwen3.6 35B-A3B: 54.22 Qwen3.6 35B-A3B: 47.47 Qwen3.6 35B-A3B: 60.92 ⚡ Qwen3.8 27B: 7.76 Qwen3.8 27B: 7.77 Qwen3.8 27B: 15.67 ⚡ Qwen3.8 27B: 13.19 ⚡ | |
| Strix Halo (Ryzen AI Max+ 395, 128 GB): Qwen3.8 27B Q6_K, 9.6 to 18 tok/s with MTP | AMD Ryzen AI Max+ 395 (Radeon 8060S iGPU) | 128 | Qwen3.8 27B: 9.59 Qwen3.8 27B: 18.01 ⚡ Qwen3.8 27B: 15.16 ⚡ | |
| DGX Spark (GB10, 128 GB): 23 models on vLLM at 128K context, Qwen3.6 35B-A3B up to 102 tok/s | Nvidia GB10 (unified) | 128 | Qwen3.6 35B-A3B: 102.3 ⚡ Qwen3.6 35B-A3B: 78.1 gpt-oss 20B: 96.2 Nemotron 3.5 Lightning 30B-A3B: 94.4 ⚡ Nemotron 3.5 Lightning 30B-A3B: 70.3 Nemotron 3 Nano Omni 30B: 58.8 Gemma 4 26B-A4B: 56.5 ⚡ Gemma 4 26B-A4B: 30.8 gpt-oss 120B: 44.4 Qwen3.8 Flash-Next: 26.8 ⚡ Nemotron 3 Super 120B-A12B: 23.0 ⚡ Nemotron 3 Super 120B-A12B: 16.1 Qwen3.8 27B: 19.0 ⚡ Qwen3.8 27B: 11.0 Llama 3.1 8B: 26.4 | |
| Minisforum MS-S1 MAX (Strix Halo, 128 GB): Qwen3.6 35B-A3B at 62 tok/s on llama.cpp Vulkan | AMD Ryzen AI Max+ 395 (Radeon 8060S iGPU) | 128 | $3,799 (≈€3,390) | Qwen3.6 35B-A3B: 62 Qwen3.6 35B-A3B: 54 Qwen3.6 35B-A3B: 51 Qwen3.6 35B-A3B: 93 ⚡ |
| DGX Spark (GB10): Qwen3.6 35B-A3B NVFP4 at 113 tok/s on vLLM with MTP | Nvidia GB10 (unified) | 128 | $3,999 (≈€3,568) | Qwen3.6 35B-A3B: 113 ⚡ Qwen3.6 35B-A3B: 86 ⚡ |
| MacBook Pro M5 Max 128 GB: Qwen3.6 35B-A3B at 271 tok/s (mlx-serve + MTP), Gemma 4 26B-A4B at 166 | Apple M5 Max (unified) | 128 | Qwen3.6 35B-A3B: 271 ⚡ Qwen3.6 35B-A3B: 240 ⚡ Qwen3.6 35B-A3B: 107 Gemma 4 26B-A4B: 166 ⚡ Qwen3.8 27B: 31 | |
| HP ZBook Ultra G1a (Strix Halo laptop, 128 GB): Qwen3.6 35B-A3B at 59.4 tok/s, gpt-oss 120B at 44.7 | AMD Ryzen AI Max+ PRO 395 (Radeon 8060S iGPU, 32 GB dedicated) | 128 | Qwen3.6 35B-A3B: 59.4 gpt-oss 120B: 44.7 Qwen3.6 35B-A3B: 54.4 Gemma 4 26B-A4B: 51.8 Qwen3.6 27B: 11.4 Gemma 4 31B: 8.8 | |
| Strix Halo mini PC (Ryzen AI Max+ 395, 128 GB): Qwen3.8 27B at 12.6 tok/s Q4_K_M, 7.7 Q8_0 | AMD Ryzen AI Max+ 395 (Radeon 8060S iGPU) | 128 | Qwen3.8 27B: 12.6 Qwen3.8 27B: 7.7 | |
| DGX Spark with Qwen3.6 27B NVFP4 + DFlash speculative decoding | Nvidia GB10 (unified) | 128 | Qwen3.6 27B: 32.83 ⚡ Qwen3.6 27B: 7.85 | |
| EPYC 7532 workstation (RTX Pro 6000 + RTX 3090 Ti, 300 W caps): Qwen3.8-Flash-Next, llama.cpp vs Strata engine | RTX Pro 6000 Blackwell Workstation (96 GB) + RTX 3090 Ti (EVGA FTW3 Ultra, 24 GB); each test used one card | 120 | Qwen3.8 Flash-Next: 102.889 ⚡ Qwen3.8 Flash-Next: 205.263 ⚡ Qwen3.8 Flash-Next: 76.337 Qwen3.8 Flash-Next: 22.657 ⚡ Qwen3.8 Flash-Next: 98.291 ⚡ | |
| 4x RTX 3090 in a 4U rack case (AM5) ⚠ unchecked | 4x Dell OEM RTX 3090 | 96 | €5,835 (≈$6,540) | |
| 4x RTX 3090 homelab on Threadripper 2920X ⚠ unchecked | 4x RTX 3090 (~$625 each, used) | 96 | $3,300 (≈€2,944) | |
| 4x RTX 3090 + 32-core Threadripper, 192 GB DDR4 ⚠ unchecked | 4x RTX 3090 | 96 | Qwen3 235B-A22B: 12-14 | |
| 4x RTX 3090 + EPYC 7443, 256 GB DDR4 8-channel ⚠ unchecked | 4x RTX 3090 | 96 | Qwen3 235B-A22B: 18 | |
| RTX Pro 6000 96GB workstation card | RTX Pro 6000 (Blackwell) | 96 | gpt-oss 120B: 196.31 gpt-oss 20B: 286.91 | |
| Strix Halo (Radeon 8060S, 96 GB): 2026 models (llamabench community) | AMD Ryzen AI Max+ (Radeon 8060S iGPU) | 96 | Qwen3.6 35B-A3B: 50.2-62.4 | |
| 4x RTX 3090 lab: Qwen3.6 on vLLM vs llama.cpp | 4x RTX 3090 | 96 | Qwen3.6 35B-A3B: 115-133 Qwen3.6 35B-A3B: 149 Qwen3.6 35B-A3B: 225-231 ⚡ | |
| RTX Pro 6000 Blackwell Max-Q (300 W cap): Qwen3.8 27B on vLLM, NVFP4 vs BF16 | RTX Pro 6000 Max-Q (Blackwell, 96 GB) | 96 | Qwen3.8 27B: 150 ⚡ Qwen3.8 27B: 85 ⚡ | |
| Mac Studio 64 GB ⚠ unchecked | Apple unified memory | 64 | $4,000 (≈€3,569) | Llama 3.3 70B: 11 |
| MacBook Pro M3 Max 40-core GPU, 64 GB | Apple M3 Max 40-core GPU | 64 | Llama 3 8B: 50.74 | |
| MacBook Pro M1 Max 32-core GPU, 64 GB | Apple M1 Max 32-core GPU | 64 | Llama 3 8B: 34.49 | |
| Apple M1 Max 64GB | Apple M1 Max (unified) | 64 | gpt-oss 20B: 75.15 | |
| Dual Tesla V100 32 GB on EPYC 7402 (bnk engine) | 2x Tesla V100 32 GB (PG500-216, sm_70) | 64 | Qwen3.8 27B: 42.4 Qwen3.8 27B: 104 ⚡ Qwen3.8 Flash-Next: 114 Qwen3.8 Flash-Next: 121 DeepSeek V4.1 Flash: 13 | |
| Dual Intel Arc Pro B70 (neural.download lab) | 2x Intel Arc Pro B70 32 GB | 64 | Qwen3.8 Flash-Next: 58.56 Qwen3.8 27B: 46.01 ⚡ | |
| Dual RTX 3090 on Ryzen 3700X (X470) ⚠ unchecked | 2x Zotac RTX 3090 | 48 | Llama 3 70B: 16.95 Llama 3 70B: 21.2 | |
| Dual used RTX 3090, ~2000 EUR total ⚠ unchecked | 2x RTX 3090 (used) | 48 | €2,000 (≈$2,242) | Llama 3.3 70B: 20 |
| 2x RTX 3090 (llm-speed signed run, Oct 2026): Llama 3.3 70B Q4_K_M at 19.1 tok/s | 2x RTX 3090 | 48 | Llama 3.3 70B: 19.1 | |
| MacBook Pro M4 Pro 48 GB: Flint 40B-A2.5B 4-bit at 127.9 tok/s (Kapitsa), 126.6 (rec), 79.6 (llama.cpp) | Apple M4 Pro (20-core GPU, unified) | 48 | Flint 40B-A2.5B: 127.9 Flint 40B-A2.5B: 126.56 Flint 40B-A2.5B: 79.57 | |
| Dual RTX PRO 4000 Blackwell on Threadripper PRO 5975WX | 2x RTX PRO 4000 Blackwell 24 GB | 48 | Qwen3.8 Flash-Next: 100 | |
| RTX 3090 + RTX 4060 Ti 16GB mixed pair ⚠ unchecked | RTX 3090 + RTX 4060 Ti 16GB | 40 | Llama 3.3 70B: 11 | |
| Apple M4 Max 36GB | Apple M4 Max (unified) | 36 | gpt-oss 20B: 92.36 | |
| Mac Studio M5 Max 36 GB: Qwen3.6 35B-A3B at 173 tok/s (Rapid-MLX), 146 on oMLX | Apple M5 Max (unified) | 36 | Qwen3.6 35B-A3B: 173.4 ⚡ Qwen3.6 35B-A3B: 145.8 Qwen3.6 35B-A3B: 124.5 Qwen3.8 27B: 37.9 ⚡ Qwen3.8 27B: 25.3 Qwen3.8 27B: 26.0 Qwen3.5 9B: 115.5 ⚡ Qwen3.5 9B: 78.0 | |
| Intel Core Ultra 265K + Radeon AI Pro R9700 | AMD Radeon AI Pro R9700 | 32 | gpt-oss 20B: 163.26 DeepSeek R1 Distill 14B: 53.72 Llama 3.2 3B: 195.91 | |
| Intel i9-13900K + RTX 5090 | RTX 5090 | 32 | Llama 3.2 3B: 271.4 | |
| M4 Mac mini (10-core) 32GB | Apple M4 (unified) | 32 | Llama 3.2 3B: 41.31 | |
| Single RTX 5090 32GB | RTX 5090 | 32 | gpt-oss 20B: 282.51 | |
| RTX 5090: 2026 models (llamabench community) | RTX 5090 | 32 | Qwen3.6 35B-A3B: 258 Qwen3.6 27B: 75.7 Gemma 4 31B: 70.0 Gemma 4 26B-A4B: 235 Gemma 4 12B: 142 gpt-oss 20B: 313 Qwen3.5 9B: 208 | |
| Radeon AI Pro R9700 (one card): Qwen3.6 35B-A3B from 4K to 256K context (AMDai lab rig) | AMD Radeon AI Pro R9700 (1 of 2 cards in the rig used) | 32 | Qwen3.6 35B-A3B: 102.39 Qwen3.6 35B-A3B: 90.52 Qwen3.6 35B-A3B: 66.43 Qwen3.6 35B-A3B: 44.4 Qwen3.6 35B-A3B: 109.73 Qwen3.6 35B-A3B: 117.69 | |
| Radeon AI Pro R9700: Qwen3.6 27B on Vulkan, HIP and Ollama, with and without MTP | AMD Radeon AI Pro R9700 (1 of 2 cards in the rig used) | 32 | Qwen3.6 27B: 66.0 ⚡ Qwen3.6 27B: 31.9 Qwen3.6 27B: 27.1 Qwen3.6 27B: 26.7 | |
| Radeon AI Pro R9700: Qwen3.8 27B on llama.cpp, Ollama and vLLM (Vadi Taslim) | AMD Radeon AI Pro R9700 (1 of 2 cards in the rig used) | 32 | Qwen3.8 27B: 57 ⚡ Qwen3.8 27B: 30 Qwen3.8 27B: 22 | |
| RTX 5090 (llm-speed signed runs, Oct 2026): gpt-oss 20B 324 tok/s, Qwen3.8 27B 75 tok/s | RTX 5090 | 32 | gpt-oss 20B: 324.2 Qwen3.6 27B: 75.1 Qwen3.8 27B: 75.2 | |
| Ryzen 9 9950X3D + RTX 5090 + 128 GB DDR5: four big MoEs with experts offloaded to RAM (Qwen3.8 Flash-Next 78.8 tok/s) | RTX 5090 | 32 | Qwen3.8 Flash-Next: 78.8 ⚡ Ornith 1.5 122B: 46.4 GLM-5.3 Flash: 30.1 ⚡ DeepSeek V4.1 Flash: 24.9 | |
| RTX 5090 workstation: Qwen3.8 Flash-Next IQ3_S at 159.2 tok/s on Strata (expert offload) | RTX 5090 | 32 | Qwen3.8 Flash-Next: 159.2 | |
| Dual RTX 5070 Ti on Ryzen 9 9900X (capy-flash maintainer rig) | 2x RTX 5070 Ti 16 GB | 32 | Qwen3.8 Flash-Next: 182 Qwen3.8 Flash-Next: 155 Qwen3.8 27B: 139 ⚡ Qwen3.8 27B: 230 ⚡ | |
| Dual RTX 5080 on EPYC 7F52 (capy-flash user report) | 2x RTX 5080 16 GB | 32 | Qwen3.8 Flash-Next: 141.3 Qwen3.8 Flash-Next: 144.4 | |
| Single Intel Arc Pro B70 32 GB (neural.download lab) | Intel Arc Pro B70 32 GB | 32 | Gemma 4 26B-A4B: 122.16 Ornith 1.5 35B-A3B: 131.46 Qwen3.8 27B: 27.8 | |
| Single Intel Arc Pro B70 with vLLM XPU cookbook | Intel Arc Pro B70 32 GB | 32 | Qwen3.8 27B: 96 ⚡ Qwen3.8 27B: 66 ⚡ Qwen3.6 35B-A3B: 120 | |
| i7-12700K + RTX 5060 Ti 16GB + RTX 3060 12GB: Qwen3.8 27B from 21 to 51 tok/s (split, MTP, single-GPU) | RTX 5060 Ti 16GB + RTX 3060 12GB | 28 | Qwen3.8 27B: 21.2 Qwen3.8 27B: 37.6 ⚡ Qwen3.8 27B: 28.0 Qwen3.8 27B: 51.4 ⚡ | |
| Single RTX 4090 (llama.cpp CUDA scoreboard) | RTX 4090 24GB | 24 | Llama 2 7B: 188.96 | |
| RTX 4090 desktop (speed comparison vs Macs) | RTX 4090 24GB | 24 | Llama 3.1 8B: 52.99 Mistral Small 22B: 26.72 | |
| EPYC + single RTX 3090 (hybrid) ⚠ unchecked | RTX 3090 | 24 | DeepSeek R1 671B: 8.97 | |
| Intel Core Ultra 265K + RTX 4090 | RTX 4090 | 24 | gpt-oss 20B: 279.65 DeepSeek R1 Distill 14B: 92.49 Llama 3.2 3B: 334.67 | |
| Ryzen 9 7900X + RTX 4090 | RTX 4090 | 24 | Llama 3.2 3B: 237.05 Llama 3.1 70B: 3.1 | |
| Single RTX 3090 24GB | RTX 3090 | 24 | gpt-oss 20B: 161.77 | |
| RTX 4090: 2026 models (llamabench community) | RTX 4090 | 24 | Qwen3.6 27B: 47.0 Gemma 4 31B: 43.8 Gemma 4 26B-A4B: 164 Gemma 4 12B: 98.1 gpt-oss 20B: 203 Qwen3.5 9B: 137 LFM2.5 24B-A2B: 261-339 Nemotron 3.5 Bolt 74B-A9B: 108 | |
| RTX 3090: 2026 models (llamabench community) | RTX 3090 | 24 | Qwen3.6 35B-A3B: 146-153 Qwen3.6 27B: 40.2 Qwen3.8 27B: 38.2-71.6 | |
| Single RTX 3090: Qwen3.6 35B-A3B (Giles Thomas) | RTX 3090 | 24 | Qwen3.6 35B-A3B: 139.6 | |
| Single RTX 3090: Qwen3.8 27B at 78-140 tok/s with the OpenJet megakernel vs llama.cpp | RTX 3090 | 24 | Qwen3.8 27B: 139.7 ⚡ Qwen3.8 27B: 72.5 ⚡ Qwen3.8 27B: 78.0 ⚡ Qwen3.8 27B: 50.9 ⚡ Qwen3.8 27B: 40.4 Qwen3.8 27B: 42.9 | |
| Radeon RX 7900 XTX (24 GB): Qwen3.6 35B-A3B up to 245K context, with MTP | AMD Radeon RX 7900 XTX | 24 | Qwen3.6 35B-A3B: 87.2 Qwen3.6 35B-A3B: 79.6 Qwen3.6 35B-A3B: 135.2 ⚡ Qwen3.6 35B-A3B: 66.0 Qwen3.6 35B-A3B: 108.2 ⚡ Qwen3.6 35B-A3B: 38.3 | |
| RTX 4090 (llm-speed signed runs, Oct 2026): Qwen3.6 35B-A3B 161 tok/s, Gemma 4 12B QAT 98 tok/s | RTX 4090 | 24 | Qwen3.6 35B-A3B: 160.6 Gemma 4 12B: 97.7 GLM 4.7 Flash: 164.9 | |
| RTX 3090 (llm-speed signed runs, Oct 2026): gpt-oss 20B 207 tok/s, Qwen3.6 27B 43 tok/s | RTX 3090 | 24 | gpt-oss 20B: 207.2 Qwen3.6 27B: 42.7 GLM 4.7 Flash: 134.6 | |
| RTX 4090 (llm-speed signed run, Oct 2026): Flint 40B-A2.5B at 169-179 tok/s on llama.cpp | RTX 4090 | 24 | Flint 40B-A2.5B: 179.4 Flint 40B-A2.5B: 169.2 | |
| RTX 3090 (llm-speed signed run, Oct 2026): Flint 40B-A2.5B Q3_K_XL at 128.9 tok/s | RTX 3090 | 24 | Flint 40B-A2.5B: 128.9 | |
| Ryzen 9 9950X + Radeon RX 7900 XT | AMD Radeon RX 7900 XT | 20 | Llama 3.2 3B: 131.2 | |
| Single RTX 4060 Ti 16GB + 64 GB RAM (partial offload) ⚠ unchecked | RTX 4060 Ti 16GB | 16 | Llama 3.3 70B: 1.5 | |
| Intel Core Ultra 265K + RTX A4000 | RTX A4000 | 16 | gpt-oss 20B: 93.55 Llama 3.2 3B: 149.44 | |
| Ryzen 9 7950X + RTX 4080 | RTX 4080 | 16 | Llama 3.2 3B: 204.45 | |
| Ryzen 9 7950X + RTX 4070 Ti Super | RTX 4070 Ti Super | 16 | Llama 3.2 3B: 198.95 | |
| Single RTX 5070 Ti 16GB | RTX 5070 Ti | 16 | gpt-oss 20B: 189.45 | |
| Single RTX 5060 Ti 16GB | RTX 5060 Ti 16GB | 16 | gpt-oss 20B: 111.51 | |
| Single RTX 5080 16GB | RTX 5080 | 16 | gpt-oss 20B: 204.85 | |
| RTX 5070 Ti: 2026 models (llamabench community) | RTX 5070 Ti | 16 | Gemma 4 12B: 87.6 | |
| RTX 4060 Ti 16GB: 2026 models (llamabench community) | RTX 4060 Ti 16GB | 16 | LFM2.5 24B-A2B: 131-150 Qwen3.8 27B: 10.7-17.6 | |
| RTX A4000 16GB (llm-speed signed run, Oct 2026): Muse Glimmer 30B UD-IQ3_XXS at 21.4 tok/s | RTX A4000 | 16 | Muse Glimmer 30B: 21.42 | |
| Intel Core Ultra 265K + RTX 3080 Ti | RTX 3080 Ti | 12 | DeepSeek R1 Distill 14B: 70.74 Llama 3.2 3B: 263.24 | |
| Intel Core Ultra 265K + RTX 4070 Ti | RTX 4070 Ti | 12 | gpt-oss 20B: 163.14 DeepSeek R1 Distill 14B: 49.74 Llama 3.2 3B: 196.15 | |
| Intel Core Ultra 265K + RTX 3060 | RTX 3060 | 12 | gpt-oss 20B: 83.96 DeepSeek R1 Distill 14B: 29.77 Llama 3.2 3B: 122.85 | |
| RTX 3060 12GB: 2026 models (llamabench community) | RTX 3060 | 12 | Gemma 4 12B: 38.5 LFM2.5 24B-A2B: 125 Qwen3.8 27B: 5.2-6.7 | |
| RTX 5070 12GB + 32 GB DDR4: Qwen3.8-Flash-Next 177B at 11.5 tok/s with expert streaming | RTX 5070 12GB | 12 | Qwen3.8 Flash-Next: 11.5 Qwen3.8 Flash-Next: 10.15 | |
| Intel Core Ultra 265K + RTX 2080 Ti | RTX 2080 Ti | 11 | DeepSeek R1 Distill 14B: 45.06 Llama 3.2 3B: 180.38 | |
| MacBook Pro M2 Max | Apple M2 Max | Llama 3.1 8B: 28.92 | ||
| Mac Studio M1 Ultra | Apple M1 Ultra | Llama 3.1 8B: 37.92 | ||
| $2K EPYC 64-core, 512 GB DDR4, CPU-only | None (CPU inference) | 0 | $2,000 (≈€1,784) | DeepSeek R1 671B: 4.31 |
| Framework Desktop (AMD Strix Halo) mainboard | AMD Ryzen AI Max+ 395 (Radeon 8060S iGPU; Ollama runs CPU only, llama.cpp runs on iGPU via Vulkan) | DeepSeek R1 Distill 8B: 19.49 DeepSeek R1 Distill 70B: 2.43 Llama 3.1 70B: 4.97 Llama 3.2 3B: 88.14 | ||
| Dell Pro Max with GB10 (Nvidia DGX Spark) | Nvidia GB10 (unified) | gpt-oss 20B: 87.97 DeepSeek R1 Distill 14B: 23.92 Llama 3.1 70B: 4.71 Llama 3.2 3B: 97.93 | ||
| AmpereOne A192-32X 512GB (CPU only) | None (CPU only) | DeepSeek R1 671B: 4.18 Llama 3.1 70B: 3.86 Llama 3.1 405B: 0.9 Llama 3.2 3B: 23.52 | ||
| 2x Mac Studio M5 Ultra cluster (Thunderbolt 5 RDMA): Muse Spark Open 816B | 2x Apple M5 Ultra (unified, Mac Studio) | Muse Spark Open 816B-A42B: 24.99 Muse Spark Open 816B-A42B: 15.74 Mistral Medium 3.5 128B: 13.69 Qwen3.8 27B: 33.27 | ||
| 4x Mac Studio M5 Ultra cluster (Thunderbolt 5 RDMA): Muse Spark Open at 36.7 tok/s | 4x Apple M5 Ultra (unified, Mac Studio) | Muse Spark Open 816B-A42B: 36.73 Mistral Medium 3.5 128B: 23.37 Qwen3.8 27B: 50.89 | ||
| Microsoft Surface Laptop Ultra (RTX Spark N1X): llama.cpp speeds on 2026 models (Micro Center) | Nvidia RTX Spark N1X, 6,144-core Blackwell GPU (unified) | Nemotron 3 Nano Omni 30B: 63.9 GLM 4.7 Flash: 55.6 Qwen3.5 122B: 35.2 Muse Glimmer 30B: 15.9 Qwen3.8 27B: 14.5 |
Weekly: new models × which rigs run them. Get the newsletter
Rent before you buy: try this model on a cloud GPU for a few dollars first: Runpod (referral link; sign up with Google to get the new-user bonus)