What GPU do you need to run Kvark Sparkle 1.1 27B?
A new Radeon RX 7900 XTX ($1,350). It is the cheapest priced graphics card that holds Kvark Sparkle 1.1 27B (about 18.3 GB at Q4) with room for about 92K tokens of context (173K with 8-bit KV cache), at about 40 tok/s estimate; that is the card price only, and it requires a separate compatible PC. 58 of the 157 GPUs and systems we track fit it.
Parameters27.78B
VRAM needed~18.3 GBfile + headroom
Single GPUs that fit58
Context window256K tokens
27.78B parameters (Hugging Face, accessed 2026-11-11). Q4_K_M GGUF: 16.81 GB (source). Architecture config (Hugging Face config.json): 16 attention layers (the rest are recurrent) × 4 KV heads × 256 head dim, so about 64 KB of FP16 KV cache per token of context (how we estimate context).
Egzakta's fine-tune of Kvark Sparkle 1.0 (itself built on Qwen3.8-27B) for document work: searching a document base, answering from what it finds, and declining when the answer isn't there. It was trained with a "stop rule" appended to the system prompt, and the card says it works in English and Serbian. Thinking is on by default. The card says the architecture is Qwen3.8-27B, unchanged; vision uses the base model's encoder and was not separately evaluated. Q4_K_M is Egzakta's official GGUF (Kvark-Sparkle-1.1-27B-Q4_K_M.gguf, 16,810,715,744 bytes); every size is in the table below. Fit leaves out the 0.93 GB vision projector, which is extra if you send images. The checkpoint and every GGUF keep a 1-layer MTP head, so speculative decoding (MTP drafting) is possible; that is not a speed figure, and the speed estimates on this page do not assume it. Release date is the Hugging Face repo creation date.
What we checked
Checked Nov 11, 2026. Official is the maker or the weight files. A vendor claim is that project's own documentation. A community claim is someone else's upload. Our check is a file or search we read. Unknown means we have no source, so it is not estimated. None of these rows is a speed we measured.
| Kind | What | Detail | Source |
|---|---|---|---|
| Official | Model card | Egzakta's card says "Architecture: Qwen3.8-27B, unchanged"; the base model is Egzakta/Kvark-Sparkle-1.0-27B, a fine-tune of Qwen3.8-27B. Context: "262,144 tokens native; evaluated up to 64k". config.json matches Qwen3.8-27B's: 64 layers (16 full attention), 4 KV heads, head size 256, the same context, 1 MTP layer and a vision tower. Hugging Face's safetensors total is 27,781,427,952 parameters (BF16). The card gives the KV cache as "64 KiB per token in BF16, 32 KiB in FP8. Only the 16 full-attention layers store it." That agrees with the 64 KB per token this page derives from config.json. | Model card, checked Nov 11, 2026 |
| Official | License | Apache 2.0. The card: "Apache 2.0, inherited from Qwen3.8-27B... Commercial use is permitted, with no revenue cap." This is a summary of the license, not legal advice. | LICENSE, checked Nov 11, 2026 |
| Official | Weight files | 13 BF16 safetensors files sum to 55,563,007,192 bytes (55.56 GB): 12 model shards and model-mtp.safetensors (849,400,424 bytes), at commit cb02a553030910573c450940eb12477d27370720. Egzakta's FP8 checkpoint for vLLM (Egzakta/Kvark-Sparkle-1.1-27B-FP8) is 30,866,867,288 bytes. These are downloads, not the memory a runtime needs. | Hugging Face API, checked Nov 11, 2026 |
| Official | GGUF files | Egzakta's own GGUF repo (commit 5cca6274422347e4dc67fd0e4cc4049aeff79e40) uses the qwen35 architecture, so stock llama.cpp loads it. The card says every file keeps the MTP head and calls Q4_K_M "recommended for 24 GB cards" (Egzakta's words). There is no bartowski repo. Every size is in the table below. | GGUF repo, checked Nov 11, 2026 |
| Vendor claim | Memory table (Egzakta's figures) | Egzakta's GGUF card lists the llama.cpp KV cache as 2 GiB at f16 or 1.1 GiB at q8_0 for 32,768 tokens; 4 or 2.1 GiB at 65,536; 8 or 4.3 GiB at 131,072; 16 or 8.5 GiB at 262,144; plus a linear-attention state under 0.2 GB at any context. Those are Egzakta's figures. The context estimates on this page come from our own formula. | GGUF card, checked Nov 11, 2026 |
| Vendor claim | MTP drafting and speed test | Egzakta's llama.cpp command uses --spec-type draft-mtp --spec-draft-n-max 3, which drafts tokens with the MTP head. That makes speculative decoding possible; it is not a speed figure. Egzakta's own test, with MTP drafting on; time per answer, not tok/s: one user, 1× H100, vLLM 0.32.0, FP8 weights, 1 MTP draft token, 20 fixed questions, MTP drafting on for both models. Median time per answer was 32.9 s for Sparkle 1.1 and 40.8 s for Qwen3.8-27B. We have not run it, and it is not a speed estimate on this page. | Model card, checked Nov 11, 2026 |
| Vendor claim | Benchmark scores | Egzakta's card reports public-benchmark scores against Qwen3.8-27B and reasoning-token counts. Those are Egzakta's own reported scores, not checked by us, and this page does not list them. | Model card, checked Nov 11, 2026 |
| Our check | stop_system.txt | The card tells users to download stop_system.txt from the repo, but at commit cb02a553 that file is not in the repo (404). The rule's full text is quoted on the card. | Hugging Face file tree, checked Nov 11, 2026 |
| Community claim | Community GGUFs | mradermacher publishes static and imatrix GGUFs (mradermacher/Kvark-Sparkle-1.1-27B-GGUF and mradermacher/Kvark-Sparkle-1.1-27B-i1-GGUF). We did not size them; this page uses Egzakta's own files. | mradermacher repo, checked Nov 11, 2026 |
| Our check | MLX | No MLX build of Kvark Sparkle 1.1 on Hugging Face as of 2026-11-11. Community MLX builds exist only for Sparkle 1.0, a different checkpoint. | Hugging Face MLX search, checked Nov 11, 2026 |
GGUF files and sizes
Exact sizes from Hugging Face, accessed 2026-11-11. The fit and context tables on this page use the Q4_K_M file (16.81 GB); a smaller file needs less memory by the difference. VRAM needed is the file + the same headroom.
| File | Published by | Bytes | Size | VRAM needed | Source |
|---|---|---|---|---|---|
| IQ4_XS | Egzakta | 15,309,040,224 | 15.31 GB | ~16.8 GB | Kvark-Sparkle-1.1-27B-IQ4_XS.gguf |
| Q4_K_M | Egzakta | 16,810,715,744 | 16.81 GB | ~18.3 GB | Kvark-Sparkle-1.1-27B-Q4_K_M.gguf |
| Q5_K_M | Egzakta | 19,535,702,624 | 19.54 GB | ~21 GB | Kvark-Sparkle-1.1-27B-Q5_K_M.gguf |
| Q6_K | Egzakta | 22,431,001,184 | 22.43 GB | ~23.9 GB | Kvark-Sparkle-1.1-27B-Q6_K.gguf |
| Q8_0 | Egzakta | 29,047,085,664 | 29.05 GB | ~30.6 GB | Kvark-Sparkle-1.1-27B-Q8_0.gguf |
| BF16 | Egzakta | 54,657,735,008 | 54.66 GB | ~56.2 GB | Kvark-Sparkle-1.1-27B-BF16.gguf |
| mmproj-F16 Vision projector, extra | Egzakta | 927,606,816 | 0.93 GB | — | mmproj-Kvark-Sparkle-1.1-27B-F16.gguf |
Single GPUs that fit, cheapest first
| Single GPU that fits | VRAM | Speed estimate | Max context estimate | Price |
|---|---|---|---|---|
| AMD Radeon RX 7900 XTX | 24 GB | ~40 tok/s | 92K · 173K (8-bit) | low $1,350 |
| NVIDIA GeForce RTX 3090 | 24 GB | ~42 tok/s | 92K · 173K (8-bit) | low $1,499 |
| Apple Mac Studio M5 Max (32-core GPU, 36 GB) | 36 GB (~24.1 usable) | ~22 tok/s | 93K · 176K (8-bit) | low $2,499 |
| NVIDIA GeForce RTX 4090 | 24 GB | ~44 tok/s | 92K · 173K (8-bit) | low $2,910 |
| NVIDIA DGX Spark (64GB) | 64 GB | ~14 tok/s | 256K (full) | low $4,999 |
| Apple Mac Studio M5 Ultra (64-core GPU, 96 GB) | 96 GB (~72 usable) | ~50 tok/s | 256K (full) | low $5,499 |
| NVIDIA GeForce RTX 5090 | 32 GB | ~70 tok/s | 211K · 256K (8-bit) | low $6,500 |
| Framework Desktop, AMD Ryzen AI Max+ PRO 495 (Strix Halo), 192 GB | 192 GB | ~12 tok/s | 256K (full) | low $6,799 |
| NVIDIA DGX Spark (128GB) | 128 GB | ~14 tok/s | 256K (full) | low $6,949.99 |
| NVIDIA RTX A6000 | 48 GB | ~35 tok/s | 256K (full) | low $8,363 |
| AMD Radeon RX 7900 XT | 20 GB | ~34 tok/s | 32K · 61K (8-bit) | No price yet |
| NVIDIA RTX 4000 Ada Generation | 20 GB | ~18 tok/s | 32K · 61K (8-bit) | No price yet |
| NVIDIA RTX 4000 SFF Ada Generation | 20 GB | ~14 tok/s | 32K · 61K (8-bit) | No price yet |
| NVIDIA RTX A4500 | 20 GB | ~30 tok/s | 32K · 61K (8-bit) | No price yet |
| Intel Arc Pro B60 | 24 GB | ~15 tok/s | 92K · 173K (8-bit) | No price yet |
| NVIDIA A10 | 24 GB | ~28 tok/s | 92K · 173K (8-bit) | No price yet |
| NVIDIA GeForce RTX 3090 Ti | 24 GB | ~44 tok/s | 92K · 173K (8-bit) | No price yet |
| NVIDIA L4 | 24 GB | ~15 tok/s | 92K · 173K (8-bit) | No price yet |
| NVIDIA Quadro RTX 6000 | 24 GB | ~31 tok/s | 92K · 173K (8-bit) | No price yet |
| NVIDIA RTX 4500 Ada Generation | 24 GB | ~21 tok/s | 92K · 173K (8-bit) | No price yet |
| NVIDIA RTX A5000 | 24 GB | ~35 tok/s | 92K · 173K (8-bit) | No price yet |
| NVIDIA RTX A5500 | 24 GB | ~35 tok/s | 92K · 173K (8-bit) | No price yet |
| NVIDIA RTX PRO 4000 Blackwell | 24 GB | ~31 tok/s | 92K · 173K (8-bit) | No price yet |
| NVIDIA RTX PRO 4000 SFF Blackwell | 24 GB | ~21 tok/s | 92K · 173K (8-bit) | No price yet |
| NVIDIA Tesla P40 | 24 GB | ~13 tok/s | 92K · 173K (8-bit) | No price yet |
| NVIDIA TITAN RTX | 24 GB | ~31 tok/s | 92K · 173K (8-bit) | No price yet |
| AMD Radeon PRO V710 | 28 GB | ~20 tok/s | 151K · 256K (8-bit) | No price yet |
| AMD Radeon AI PRO R9600 | 32 GB | ~27 tok/s | 211K · 256K (8-bit) | No price yet |
| AMD Radeon AI PRO R9600D | 32 GB | ~27 tok/s | 211K · 256K (8-bit) | No price yet |
| AMD Radeon AI PRO R9700 | 32 GB | ~27 tok/s | 211K · 256K (8-bit) | No price yet |
| AMD Radeon AI PRO R9700S | 32 GB | ~27 tok/s | 211K · 256K (8-bit) | No price yet |
| AMD Radeon PRO V620 | 32 GB | ~22 tok/s | 211K · 256K (8-bit) | No price yet |
| AMD Radeon PRO W6800 | 32 GB | ~22 tok/s | 211K · 256K (8-bit) | No price yet |
| AMD Radeon PRO W7800 32GB | 32 GB | ~25 tok/s | 211K · 256K (8-bit) | No price yet |
| Intel Arc Pro B70 | 32 GB | ~20 tok/s | 211K · 256K (8-bit) | No price yet |
| NVIDIA Quadro GV100 | 32 GB | ~39 tok/s | 211K · 256K (8-bit) | No price yet |
| NVIDIA RTX 5000 Ada Generation | 32 GB | ~27 tok/s | 211K · 256K (8-bit) | No price yet |
| NVIDIA RTX PRO 4500 Blackwell | 32 GB | ~40 tok/s | 211K · 256K (8-bit) | No price yet |
| NVIDIA V100 PCIe 32GB | 32 GB | ~40 tok/s | 211K · 256K (8-bit) | No price yet |
| NVIDIA V100S PCIe 32GB | 32 GB | ~49 tok/s | 211K · 256K (8-bit) | No price yet |
| Apple M5 Pro 48GB | 48 GB (~36 usable) | ~15 tok/s | 256K (full) | No price yet |
| NVIDIA A100 40GB PCIe | 40 GB | ~63 tok/s | 256K (full) | No price yet |
| NVIDIA A800 40GB Active | 40 GB | ~63 tok/s | 256K (full) | No price yet |
| AMD Radeon PRO W7800 48GB | 48 GB | ~36 tok/s | 256K (full) | No price yet |
| AMD Radeon PRO W7900 | 48 GB | ~36 tok/s | 256K (full) | No price yet |
| AMD Radeon PRO W7900 Dual Slot | 48 GB | ~36 tok/s | 256K (full) | No price yet |
| Apple Mac Studio M5 Max (40-core GPU, 64 GB) | 64 GB (~48 usable) | ~29 tok/s | 256K (full) | No price yet |
| NVIDIA A40 | 48 GB | ~32 tok/s | 256K (full) | No price yet |
| NVIDIA L40 | 48 GB | ~39 tok/s | 256K (full) | No price yet |
| NVIDIA L40S | 48 GB | ~39 tok/s | 256K (full) | No price yet |
| NVIDIA Quadro RTX 8000 | 48 GB | ~31 tok/s | 256K (full) | No price yet |
| NVIDIA RTX 6000 Ada Generation | 48 GB | ~43 tok/s | 256K (full) | No price yet |
| NVIDIA RTX PRO 5000 Blackwell 48GB | 48 GB | ~56 tok/s | 256K (full) | No price yet |
| NVIDIA RTX PRO 5000 Blackwell 72GB | 72 GB | ~56 tok/s | 256K (full) | No price yet |
| NVIDIA A100 80GB PCIe | 80 GB | ~74 tok/s | 256K (full) | No price yet |
| NVIDIA RTX PRO 6000 Blackwell Max-Q | 96 GB | ~70 tok/s | 256K (full) | No price yet |
| NVIDIA RTX PRO 6000 Blackwell Server Edition | 96 GB | ~65 tok/s | 256K (full) | No price yet |
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition | 96 GB | ~70 tok/s | 256K (full) | No price yet |
Every device that runs it
Graphics cards, Macs and mini PCs from the wider hardware list behind My rig, smallest usable memory first. Not sure about yours? Check your exact rig.
Show all 108 devices that run it
| Device | Type | Memory | Speed estimate | Max context estimate | Price |
|---|---|---|---|---|---|
| AMD Radeon RX 7900 XT | GPU | 20 GB | ~34 tok/s | 32K · 61K (8-bit) | No price yet |
| NVIDIA RTX 4000 Ada Generation | Workstation GPU | 20 GB | ~18 tok/s | 32K · 61K (8-bit) | No price yet |
| NVIDIA RTX 4000 SFF Ada Generation | Workstation GPU | 20 GB | ~14 tok/s | 32K · 61K (8-bit) | No price yet |
| NVIDIA RTX A4500 | Workstation GPU | 20 GB | ~30 tok/s | 32K · 61K (8-bit) | No price yet |
| Apple M1 Max 32GB | Mac | 32 GB (~21.4 usable) | ~19 tok/s | 53K · 100K (8-bit) | No price yet |
| Apple M1 Pro 32GB | Mac | 32 GB (~21.4 usable) | ~10 tok/s | 53K · 100K (8-bit) | No price yet |
| Apple M2 Max (30-core GPU) 32GB | Mac | 32 GB (~21.4 usable) | ~19 tok/s | 53K · 100K (8-bit) | No price yet |
| Apple M2 Max (38-core GPU) 32GB | Mac | 32 GB (~21.4 usable) | ~19 tok/s | 53K · 100K (8-bit) | No price yet |
| Apple M2 Pro 32GB | Mac | 32 GB (~21.4 usable) | ~10 tok/s | 53K · 100K (8-bit) | No price yet |
| Apple M4 32GB | Mac | 32 GB (~21.4 usable) | ~6 tok/s | 53K · 100K (8-bit) | No price yet |
| Apple M5 32GB | Mac | 32 GB (~21.4 usable) | ~8 tok/s | 53K · 100K (8-bit) | No price yet |
| Apple M6 32GB | Mac | 32 GB (~21.4 usable) | ~9 tok/s | 53K · 100K (8-bit) | No price yet |
| AMD Radeon RX 7900 XTX | GPU | 24 GB | ~40 tok/s | 92K · 173K (8-bit) | low $1,350 |
| Intel Arc Pro B60 | Workstation GPU | 24 GB | ~15 tok/s | 92K · 173K (8-bit) | No price yet |
| NVIDIA A10 | Data center GPU | 24 GB | ~28 tok/s | 92K · 173K (8-bit) | No price yet |
| NVIDIA GeForce RTX 3090 | GPU | 24 GB | ~42 tok/s | 92K · 173K (8-bit) | low $1,499 |
| NVIDIA GeForce RTX 3090 Ti | GPU | 24 GB | ~44 tok/s | 92K · 173K (8-bit) | No price yet |
| NVIDIA GeForce RTX 4090 | GPU | 24 GB | ~44 tok/s | 92K · 173K (8-bit) | low $2,910 |
| NVIDIA L4 | Data center GPU | 24 GB | ~15 tok/s | 92K · 173K (8-bit) | No price yet |
| NVIDIA Quadro RTX 6000 | Workstation GPU | 24 GB | ~31 tok/s | 92K · 173K (8-bit) | No price yet |
| NVIDIA RTX 4500 Ada Generation | Workstation GPU | 24 GB | ~21 tok/s | 92K · 173K (8-bit) | No price yet |
| NVIDIA RTX A5000 | Workstation GPU | 24 GB | ~35 tok/s | 92K · 173K (8-bit) | No price yet |
| NVIDIA RTX A5500 | Workstation GPU | 24 GB | ~35 tok/s | 92K · 173K (8-bit) | No price yet |
| NVIDIA RTX PRO 4000 Blackwell | Workstation GPU | 24 GB | ~31 tok/s | 92K · 173K (8-bit) | No price yet |
| NVIDIA RTX PRO 4000 SFF Blackwell | Workstation GPU | 24 GB | ~21 tok/s | 92K · 173K (8-bit) | No price yet |
| NVIDIA RTX Spark N1X laptop (6144-core GPU, 24GB) | Laptop | 24 GB | ~15 tok/s | 92K · 173K (8-bit) | No price yet |
| NVIDIA Tesla P40 | Data center GPU | 24 GB | ~13 tok/s | 92K · 173K (8-bit) | No price yet |
| NVIDIA TITAN RTX | GPU | 24 GB | ~31 tok/s | 92K · 173K (8-bit) | No price yet |
| Apple M3 Max (30-core GPU) 36GB | Mac | 36 GB (~24.1 usable) | ~15 tok/s | 93K · 176K (8-bit) | No price yet |
| Apple M3 Pro 36GB | Mac | 36 GB (~24.1 usable) | ~8 tok/s | 93K · 176K (8-bit) | No price yet |
| Apple M4 Max (32-core GPU) 36GB | Mac | 36 GB (~24.1 usable) | ~20 tok/s | 93K · 176K (8-bit) | No price yet |
| Apple Mac Studio M5 Max (32-core GPU, 36 GB) | Mac | 36 GB (~24.1 usable) | ~22 tok/s | 93K · 176K (8-bit) | low $2,499 |
| AMD Radeon PRO V710 | Data center GPU | 28 GB | ~20 tok/s | 151K · 256K (8-bit) | No price yet |
| AMD Radeon AI PRO R9600 | Workstation GPU | 32 GB | ~27 tok/s | 211K · 256K (8-bit) | No price yet |
| AMD Radeon AI PRO R9600D | Workstation GPU | 32 GB | ~27 tok/s | 211K · 256K (8-bit) | No price yet |
| AMD Radeon AI PRO R9700 | Workstation GPU | 32 GB | ~27 tok/s | 211K · 256K (8-bit) | No price yet |
| AMD Radeon AI PRO R9700S | Workstation GPU | 32 GB | ~27 tok/s | 211K · 256K (8-bit) | No price yet |
| AMD Radeon PRO V620 | Data center GPU | 32 GB | ~22 tok/s | 211K · 256K (8-bit) | No price yet |
| AMD Radeon PRO W6800 | Workstation GPU | 32 GB | ~22 tok/s | 211K · 256K (8-bit) | No price yet |
| AMD Radeon PRO W7800 32GB | Workstation GPU | 32 GB | ~25 tok/s | 211K · 256K (8-bit) | No price yet |
| AMD Ryzen AI Max 385 (32GB) | Mini PC | 32 GB | ~11 tok/s | 211K · 256K (8-bit) | No price yet |
| AMD Ryzen AI Max 390 (32GB) | Mini PC | 32 GB | ~11 tok/s | 211K · 256K (8-bit) | No price yet |
| AMD Ryzen AI Max+ 395 (32GB) | Mini PC | 32 GB | ~11 tok/s | 211K · 256K (8-bit) | No price yet |
| Intel Arc Pro B70 | Workstation GPU | 32 GB | ~20 tok/s | 211K · 256K (8-bit) | No price yet |
| NVIDIA GeForce RTX 5090 | GPU | 32 GB | ~70 tok/s | 211K · 256K (8-bit) | low $6,500 |
| NVIDIA Quadro GV100 | Workstation GPU | 32 GB | ~39 tok/s | 211K · 256K (8-bit) | No price yet |
| NVIDIA RTX 5000 Ada Generation | Workstation GPU | 32 GB | ~27 tok/s | 211K · 256K (8-bit) | No price yet |
| NVIDIA RTX PRO 4500 Blackwell | Workstation GPU | 32 GB | ~40 tok/s | 211K · 256K (8-bit) | No price yet |
| NVIDIA RTX Spark N1X laptop (6144-core GPU, 32GB) | Laptop | 32 GB | ~15 tok/s | 211K · 256K (8-bit) | No price yet |
| NVIDIA V100 PCIe 32GB | Data center GPU | 32 GB | ~40 tok/s | 211K · 256K (8-bit) | No price yet |
| NVIDIA V100S PCIe 32GB | Data center GPU | 32 GB | ~49 tok/s | 211K · 256K (8-bit) | No price yet |
| Apple M3 Max (40-core GPU) 48GB | Mac | 48 GB (~36 usable) | ~19 tok/s | 256K (full) | No price yet |
| Apple M4 Max (40-core GPU) 48GB | Mac | 48 GB (~36 usable) | ~26 tok/s | 256K (full) | No price yet |
| Apple M4 Pro 48GB | Mac | 48 GB (~36 usable) | ~14 tok/s | 256K (full) | No price yet |
| Apple M5 Max (40-core GPU) 48GB | Mac | 48 GB (~36 usable) | ~29 tok/s | 256K (full) | No price yet |
| Apple M5 Pro 48GB | Mac | 48 GB (~36 usable) | ~15 tok/s | 256K (full) | No price yet |
| NVIDIA A100 40GB PCIe | Data center GPU | 40 GB | ~63 tok/s | 256K (full) | No price yet |
| NVIDIA A800 40GB Active | Workstation GPU | 40 GB | ~63 tok/s | 256K (full) | No price yet |
| AMD Radeon PRO W7800 48GB | Workstation GPU | 48 GB | ~36 tok/s | 256K (full) | No price yet |
| AMD Radeon PRO W7900 | Workstation GPU | 48 GB | ~36 tok/s | 256K (full) | No price yet |
| AMD Radeon PRO W7900 Dual Slot | Workstation GPU | 48 GB | ~36 tok/s | 256K (full) | No price yet |
| Apple M1 Max 64GB | Mac | 64 GB (~48 usable) | ~19 tok/s | 256K (full) | No price yet |
| Apple M1 Ultra 64GB | Mac | 64 GB (~48 usable) | ~36 tok/s | 256K (full) | No price yet |
| Apple M2 Max (30-core GPU) 64GB | Mac | 64 GB (~48 usable) | ~19 tok/s | 256K (full) | No price yet |
| Apple M2 Max (38-core GPU) 64GB | Mac | 64 GB (~48 usable) | ~19 tok/s | 256K (full) | No price yet |
| Apple M2 Ultra 64GB | Mac | 64 GB (~48 usable) | ~36 tok/s | 256K (full) | No price yet |
| Apple M3 Max (40-core GPU) 64GB | Mac | 64 GB (~48 usable) | ~19 tok/s | 256K (full) | No price yet |
| Apple M4 Max (40-core GPU) 64GB | Mac | 64 GB (~48 usable) | ~26 tok/s | 256K (full) | No price yet |
| Apple M4 Pro 64GB | Mac | 64 GB (~48 usable) | ~14 tok/s | 256K (full) | No price yet |
| Apple M5 Pro 64GB | Mac | 64 GB (~48 usable) | ~15 tok/s | 256K (full) | No price yet |
| Apple Mac Studio M5 Max (40-core GPU, 64 GB) | Mac | 64 GB (~48 usable) | ~29 tok/s | 256K (full) | No price yet |
| NVIDIA A40 | Data center GPU | 48 GB | ~32 tok/s | 256K (full) | No price yet |
| NVIDIA L40 | Data center GPU | 48 GB | ~39 tok/s | 256K (full) | No price yet |
| NVIDIA L40S | Data center GPU | 48 GB | ~39 tok/s | 256K (full) | No price yet |
| NVIDIA Quadro RTX 8000 | Workstation GPU | 48 GB | ~31 tok/s | 256K (full) | No price yet |
| NVIDIA RTX 6000 Ada Generation | Workstation GPU | 48 GB | ~43 tok/s | 256K (full) | No price yet |
| NVIDIA RTX A6000 | Workstation GPU | 48 GB | ~35 tok/s | 256K (full) | low $8,363 |
| NVIDIA RTX PRO 5000 Blackwell 48GB | Workstation GPU | 48 GB | ~56 tok/s | 256K (full) | No price yet |
| NVIDIA RTX Spark N1X laptop (6144-core GPU, 48GB) | Laptop | 48 GB | ~15 tok/s | 256K (full) | No price yet |
| AMD Ryzen AI Max 385 (64GB) | Mini PC | 64 GB | ~11 tok/s | 256K (full) | No price yet |
| AMD Ryzen AI Max 390 (64GB) | Mini PC | 64 GB | ~11 tok/s | 256K (full) | No price yet |
| AMD Ryzen AI Max+ 395 (64GB) | Mini PC | 64 GB | ~11 tok/s | 256K (full) | low $2,199.99 |
| NVIDIA DGX Spark (64GB) | Mini PC | 64 GB | ~14 tok/s | 256K (full) | low $4,999 |
| NVIDIA RTX Spark N1X laptop (6144-core GPU, 64GB) | Laptop | 64 GB | ~15 tok/s | 256K (full) | No price yet |
| Apple M2 Max (38-core GPU) 96GB | Mac | 96 GB (~72 usable) | ~19 tok/s | 256K (full) | No price yet |
| Apple M3 Max (30-core GPU) 96GB | Mac | 96 GB (~72 usable) | ~15 tok/s | 256K (full) | No price yet |
| Apple M3 Ultra 96GB | Mac | 96 GB (~72 usable) | ~37 tok/s | 256K (full) | No price yet |
| Apple Mac Studio M5 Ultra (64-core GPU, 96 GB) | Mac | 96 GB (~72 usable) | ~50 tok/s | 256K (full) | low $5,499 |
| NVIDIA RTX PRO 5000 Blackwell 72GB | Workstation GPU | 72 GB | ~56 tok/s | 256K (full) | No price yet |
| NVIDIA A100 80GB PCIe | Data center GPU | 80 GB | ~74 tok/s | 256K (full) | No price yet |
| Apple M1 Ultra 128GB | Mac | 128 GB (~96 usable) | ~36 tok/s | 256K (full) | No price yet |
| Apple M2 Ultra 128GB | Mac | 128 GB (~96 usable) | ~36 tok/s | 256K (full) | No price yet |
| Apple M3 Max (40-core GPU) 128GB | Mac | 128 GB (~96 usable) | ~19 tok/s | 256K (full) | No price yet |
| Apple M4 Max (40-core GPU) 128GB | Mac | 128 GB (~96 usable) | ~26 tok/s | 256K (full) | No price yet |
| Apple M5 Max (40-core GPU) 128GB | Mac | 128 GB (~96 usable) | ~29 tok/s | 256K (full) | No price yet |
| NVIDIA RTX PRO 6000 Blackwell Max-Q | Workstation GPU | 96 GB | ~70 tok/s | 256K (full) | No price yet |
| NVIDIA RTX PRO 6000 Blackwell Server Edition | Data center GPU | 96 GB | ~65 tok/s | 256K (full) | No price yet |
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition | Workstation GPU | 96 GB | ~70 tok/s | 256K (full) | No price yet |
| AMD Ryzen AI Max 385 (128GB) | Mini PC | 128 GB | ~11 tok/s | 256K (full) | No price yet |
| AMD Ryzen AI Max 390 (128GB) | Mini PC | 128 GB | ~11 tok/s | 256K (full) | No price yet |
| AMD Ryzen AI Max+ 395 (128GB) | Mini PC | 128 GB | ~11 tok/s | 256K (full) | low $3,499.99 |
| NVIDIA DGX Spark (128GB) | Mini PC | 128 GB | ~14 tok/s | 256K (full) | low $6,949.99 |
| NVIDIA RTX Spark N1X laptop (6144-core GPU, 128GB) | Laptop | 128 GB | ~15 tok/s | 256K (full) | No price yet |
| Apple M2 Ultra 192GB | Mac | 192 GB (~144 usable) | ~36 tok/s | 256K (full) | No price yet |
| Apple M3 Ultra 256GB | Mac | 256 GB (~192 usable) | ~37 tok/s | 256K (full) | No price yet |
| Apple M5 Ultra 256GB | Mac | 256 GB (~192 usable) | ~50 tok/s | 256K (full) | No price yet |
| Framework Desktop, AMD Ryzen AI Max+ PRO 495 (Strix Halo), 192 GB | Mini PC | 192 GB | ~12 tok/s | 256K (full) | low $6,799 |
| Apple M5 Ultra 512GB | Mac | 512 GB (~384 usable) | ~50 tok/s | 256K (full) | No price yet |
Multi-GPU: cards that need two or more
| Cards | Total VRAM | Speed estimate | Max context estimate |
|---|---|---|---|
| 2 × Intel Arc B580 | 24 GB | ~15 tok/s | 92K · 173K (8-bit) |
| 2 × NVIDIA GeForce RTX 3060 12 GB | 24 GB | ~18 tok/s | 92K · 173K (8-bit) |
| 2 × NVIDIA GeForce RTX 4070 SUPER | 24 GB | ~24 tok/s | 92K · 173K (8-bit) |
| 3 × NVIDIA GeForce RTX 5060 | 24 GB | ~22 tok/s | 92K · 173K (8-bit) |
| 3 × NVIDIA GeForce RTX 5060 Ti 8 GB | 24 GB | ~22 tok/s | 92K · 173K (8-bit) |
| 2 × NVIDIA GeForce RTX 5070 | 24 GB | ~31 tok/s | 92K · 173K (8-bit) |
| 2 × AMD Radeon RX 7800 XT | 32 GB | ~27 tok/s | 211K · 256K (8-bit) |
| 2 × AMD Radeon RX 9060 XT 16 GB | 32 GB | ~14 tok/s | 211K · 256K (8-bit) |
| 2 × AMD Radeon RX 9070 | 32 GB | ~27 tok/s | 211K · 256K (8-bit) |
| 2 × AMD Radeon RX 9070 XT | 32 GB | ~27 tok/s | 211K · 256K (8-bit) |
| 2 × NVIDIA GeForce RTX 5060 Ti 16 GB | 32 GB | ~22 tok/s | 211K · 256K (8-bit) |
| 2 × NVIDIA GeForce RTX 5070 Ti | 32 GB | ~40 tok/s | 211K · 256K (8-bit) |
| 2 × NVIDIA GeForce RTX 5080 | 32 GB | ~43 tok/s | 211K · 256K (8-bit) |
Speed calculator
estimateEstimate = bandwidth ÷ GB read per token (the model's listed GGUF file: Q4_K_M for most models, otherwise the quant named in the list; scaled by active/total parameters for MoE), scaled by a per-platform efficiency and fixed per-token costs calibrated to 45 measured llama.cpp benchmarks (2025–26). Fit = file size + 1.5 GB headroom ≤ total VRAM (on Macs, the part of unified memory macOS lets the GPU use by default), and room for at least 4K tokens of context with an 8-bit KV cache. For a model with no KV cache data, only the file and headroom are checked, so a fit says nothing about context. Multi-GPU: VRAM adds, speed stays one card's (a layer split runs the cards in turn); max context uses the total. Speed is for short context. Max context = what is left after the file and a 1 GB runtime reserve ÷ KV cache bytes per token, default FP16 and 8-bit (q8_0) cache. How speed works · how context works