What GPU do you need to run PLaMo 3 Translate 31B?
A new Radeon RX 7900 XTX ($1,350). It is the cheapest priced graphics card that holds PLaMo 3 Translate 31B (about 21 GB at Q4) with room for its full 32K context, at about 38 tok/s estimate; that is the card price only, and it requires a separate compatible PC. 54 of the 157 GPUs and systems we track fit it.
Translation only PLaMo 3 Translate 31B is a translation model, not a general chat assistant. The fits and speeds below say whether it runs on a machine, not that it is a good everyday chat model, so the site's automatic picks (the largest model a card runs, can-run pages) leave it out. License: PLaMo Community License.
Parameters31.76B
VRAM needed~21 GBfile + headroom
Single GPUs that fit54
Context window32K tokens
31.76B parameters (Hugging Face, accessed 2026-11-10). Q4_K_M GGUF: 19.53 GB (source). Architecture config (Hugging Face config.json; 64 layers, window_size 2048, sliding_window_pattern 8: llama.cpp plamo3 (load_swa_pattern(ml, 8), LLAMA_SWA_TYPE_STANDARD) makes layer il sliding-window unless il % 8 == 7, so 8 full-attention + 56 sliding-window layers; the official GGUF metadata says the same (sliding_window 2048, sliding_window_pattern 8). max_ctx is the model card's 32,768-token training limit for input and output combined (the context field), not config.json's max_position_embeddings 262,144): 8 global attention layers × 8 KV heads × 128 head dim, plus 56 sliding-window layers (2048-token window), so about 32 KB of FP16 KV cache per token of context (how we estimate context).
Translation model, not a general chat assistant: it takes text in one language and returns a translation, through its own translation prompt format. Dense model, one of four PLaMo 3 Translate sizes. Context is 32,768 tokens because the model card says training used at most 32,768 tokens of input and output combined and asks you to split anything longer; config.json allows 262,144 positions, which this page does not treat as usable. Commercial use needs advance registration with Preferred Networks (see License below). Release date is the date of PFN's release blog post.
What we checked
Checked Nov 10, 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 | PFN's card describes PLaMo 3 Translate as a translation-specialized model, tagged translation on Hugging Face, not a general chat or coding assistant. It uses a translation chat template (the PFML3 prompt format). The release blog covers four sizes (0.8B, 2B, 8B, 31B) and 53 languages, and says the model accepts surrounding context and glossaries and can write plain or polite Japanese. The card says the longest training sequence was 32,768 tokens for input and output combined and to split anything longer, so this page uses 32,768 as the context, not config.json's 262,144. Safetensors total 31,757,456,384 parameters (31.76B), BF16, dense. The repo is not gated. | Model card, checked Nov 10, 2026 |
| Official | GGUF files | PFN publishes the GGUF build itself (pfnet/plamo-3-translate-31b-GGUF), one file per quant: Q4_K_M 19,526,593,952 bytes, Q3_K_M 15,863,131,552 bytes, Q2_K 12,609,924,512 bytes. This page's fit and speed estimates use the Q4_K_M file. The 31B card's example runs llama-server with -c 16384, --jinja and --no-prefill-assistant; 16,384 is only their example. We read the sizes from the Hugging Face file list and did not download them. | GGUF repo, checked Nov 10, 2026 |
| Official | MLX 4-bit build | PFN also publishes an MLX 4-bit build for Apple silicon (pfnet/plamo-3-translate-31b-MLX-4bit), 18,656,969,144 bytes of safetensors. Its card says to run it with mlx-lm with trust_remote_code=True. Other official formats in the release are BF16, FP8, W4A16, W4AFP8 and NVFP4. We have not run the MLX build on a Mac, and no MLX speed is listed. | MLX-4bit repo, checked Nov 10, 2026 |
| Official | License: commercial use needs registration | PLaMo Community License. Personal and non-commercial use, including academic use, is free with no registration. Commercial use is free only if you (with your affiliates) had revenue of JPY 1 billion or less in the most recent fiscal year AND you register with PFN in advance through its form (https://forms.gle/mTL8tBLrMYXKNZD56). Above JPY 1 billion you need a separate license agreement with PFN (https://translate.preferredai.jp/llm/contact/). This is a summary of the license, not legal advice. | PLaMo Community License, checked Nov 10, 2026 |
| Vendor claim | PFN's benchmark claims | PFN's own claim: even the 0.8B model outperforms existing translation-specialized models (TranslateGemma, Hy-MT2) on Japanese-English and on Japanese to and from other languages, and PFN calls PLaMo 3 Translate the world's best translation-specialized LLM. Its benchmarks are PFTB and SaaS (PFN internal sets), WMT24++ (COMET-22) and FLORES+ (chrF++). PFN's card also says the Q2_K GGUF loses accuracy on some metrics while the other formats stay close to BF16. These are PFN's results; we have not reproduced any of them. | PFN release blog, checked Nov 10, 2026 |
| Our check | llama.cpp and mlx-lm support | llama.cpp master has a plamo3 architecture (src/models/plamo3.cpp), and the official GGUF header says general.architecture plamo3. With a sliding window set, llama.cpp loads a sliding-window pattern of 8: in each group of 8 layers, 7 use a 2,048-token sliding window and 1 uses full attention. For this size that is 8 full-attention and 56 sliding-window layers of 64, with 8 KV heads of size 128, which is what the context estimates use. mlx-lm master has mlx_lm/models/plamo3.py. | llama.cpp plamo3.cpp, checked Nov 10, 2026 |
| Unknown | Not tested by us | We have not run any PLaMo 3 Translate model. No speed on this page is measured: tok/s figures are the site's bandwidth estimates. Translation quality is not tested by us, and the sizes are not compared with each other or with other models here. | Model card does not give this, checked Nov 10, 2026 |
GGUF files and sizes
Exact sizes from Hugging Face, accessed 2026-11-10. The fit and context tables on this page use the Q4_K_M file (19.53 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 |
|---|---|---|---|---|---|
| Q2_K | Preferred Networks | 12,609,924,512 | 12.61 GB | ~14.1 GB | Hugging Face |
| Q3_K_M | Preferred Networks | 15,863,131,552 | 15.86 GB | ~17.4 GB | Hugging Face |
| Q4_K_M | Preferred Networks | 19,526,593,952 | 19.53 GB | ~21 GB | plamo-3-translate-31b-Q4_K_M.gguf |
Other places to get it
Checked Nov 14, 2026. We list only copies we opened and checked. Check file hashes for copies from other uploaders.
- ModelScope: Other uploaderSame account name as on Hugging Face; we found no link from the lab to it
Single GPUs that fit, cheapest first
| Single GPU that fits | VRAM | Speed estimate | Max context estimate | Price |
|---|---|---|---|---|
| AMD Radeon RX 7900 XTX | 24 GB | ~38 tok/s | 32K (full) | low $1,350 |
| NVIDIA GeForce RTX 3090 | 24 GB | ~40 tok/s | 32K (full) | low $1,499 |
| Apple Mac Studio M5 Max (32-core GPU, 36 GB) | 36 GB (~24.1 usable) | ~20 tok/s | 32K (full) | low $2,499 |
| NVIDIA GeForce RTX 4090 | 24 GB | ~43 tok/s | 32K (full) | low $2,910 |
| NVIDIA DGX Spark (64GB) | 64 GB | ~12 tok/s | 32K (full) | low $4,999 |
| Apple Mac Studio M5 Ultra (64-core GPU, 96 GB) | 96 GB (~72 usable) | ~50 tok/s | 32K (full) | low $5,499 |
| NVIDIA GeForce RTX 5090 | 32 GB | ~74 tok/s | 32K (full) | low $6,500 |
| Framework Desktop, AMD Ryzen AI Max+ PRO 495 (Strix Halo), 192 GB | 192 GB | ~11 tok/s | 32K (full) | low $6,799 |
| NVIDIA DGX Spark (128GB) | 128 GB | ~12 tok/s | 32K (full) | low $6,949.99 |
| NVIDIA RTX A6000 | 48 GB | ~33 tok/s | 32K (full) | low $8,363 |
| Intel Arc Pro B60 | 24 GB | ~14 tok/s | 32K (full) | No price yet |
| NVIDIA A10 | 24 GB | ~26 tok/s | 32K (full) | No price yet |
| NVIDIA GeForce RTX 3090 Ti | 24 GB | ~43 tok/s | 32K (full) | No price yet |
| NVIDIA L4 | 24 GB | ~13 tok/s | 32K (full) | No price yet |
| NVIDIA Quadro RTX 6000 | 24 GB | ~29 tok/s | 32K (full) | No price yet |
| NVIDIA RTX 4500 Ada Generation | 24 GB | ~19 tok/s | 32K (full) | No price yet |
| NVIDIA RTX A5000 | 24 GB | ~33 tok/s | 32K (full) | No price yet |
| NVIDIA RTX A5500 | 24 GB | ~33 tok/s | 32K (full) | No price yet |
| NVIDIA RTX PRO 4000 Blackwell | 24 GB | ~29 tok/s | 32K (full) | No price yet |
| NVIDIA RTX PRO 4000 SFF Blackwell | 24 GB | ~19 tok/s | 32K (full) | No price yet |
| NVIDIA Tesla P40 | 24 GB | ~11 tok/s | 32K (full) | No price yet |
| NVIDIA TITAN RTX | 24 GB | ~29 tok/s | 32K (full) | No price yet |
| AMD Radeon PRO V710 | 28 GB | ~18 tok/s | 32K (full) | No price yet |
| AMD Radeon AI PRO R9600 | 32 GB | ~25 tok/s | 32K (full) | No price yet |
| AMD Radeon AI PRO R9600D | 32 GB | ~25 tok/s | 32K (full) | No price yet |
| AMD Radeon AI PRO R9700 | 32 GB | ~25 tok/s | 32K (full) | No price yet |
| AMD Radeon AI PRO R9700S | 32 GB | ~25 tok/s | 32K (full) | No price yet |
| AMD Radeon PRO V620 | 32 GB | ~20 tok/s | 32K (full) | No price yet |
| AMD Radeon PRO W6800 | 32 GB | ~20 tok/s | 32K (full) | No price yet |
| AMD Radeon PRO W7800 32GB | 32 GB | ~23 tok/s | 32K (full) | No price yet |
| Intel Arc Pro B70 | 32 GB | ~18 tok/s | 32K (full) | No price yet |
| NVIDIA Quadro GV100 | 32 GB | ~37 tok/s | 32K (full) | No price yet |
| NVIDIA RTX 5000 Ada Generation | 32 GB | ~25 tok/s | 32K (full) | No price yet |
| NVIDIA RTX PRO 4500 Blackwell | 32 GB | ~39 tok/s | 32K (full) | No price yet |
| NVIDIA V100 PCIe 32GB | 32 GB | ~39 tok/s | 32K (full) | No price yet |
| NVIDIA V100S PCIe 32GB | 32 GB | ~48 tok/s | 32K (full) | No price yet |
| Apple M5 Pro 48GB | 48 GB (~36 usable) | ~14 tok/s | 32K (full) | No price yet |
| NVIDIA A100 40GB PCIe | 40 GB | ~65 tok/s | 32K (full) | No price yet |
| NVIDIA A800 40GB Active | 40 GB | ~65 tok/s | 32K (full) | No price yet |
| AMD Radeon PRO W7800 48GB | 48 GB | ~34 tok/s | 32K (full) | No price yet |
| AMD Radeon PRO W7900 | 48 GB | ~34 tok/s | 32K (full) | No price yet |
| AMD Radeon PRO W7900 Dual Slot | 48 GB | ~34 tok/s | 32K (full) | No price yet |
| Apple Mac Studio M5 Max (40-core GPU, 64 GB) | 64 GB (~48 usable) | ~27 tok/s | 32K (full) | No price yet |
| NVIDIA A40 | 48 GB | ~30 tok/s | 32K (full) | No price yet |
| NVIDIA L40 | 48 GB | ~37 tok/s | 32K (full) | No price yet |
| NVIDIA L40S | 48 GB | ~37 tok/s | 32K (full) | No price yet |
| NVIDIA Quadro RTX 8000 | 48 GB | ~29 tok/s | 32K (full) | No price yet |
| NVIDIA RTX 6000 Ada Generation | 48 GB | ~41 tok/s | 32K (full) | No price yet |
| NVIDIA RTX PRO 5000 Blackwell 48GB | 48 GB | ~56 tok/s | 32K (full) | No price yet |
| NVIDIA RTX PRO 5000 Blackwell 72GB | 72 GB | ~56 tok/s | 32K (full) | No price yet |
| NVIDIA A100 80GB PCIe | 80 GB | ~79 tok/s | 32K (full) | No price yet |
| NVIDIA RTX PRO 6000 Blackwell Max-Q | 96 GB | ~74 tok/s | 32K (full) | No price yet |
| NVIDIA RTX PRO 6000 Blackwell Server Edition | 96 GB | ~67 tok/s | 32K (full) | No price yet |
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition | 96 GB | ~74 tok/s | 32K (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 104 devices that run it
| Device | Type | Memory | Speed estimate | Max context estimate | Price |
|---|---|---|---|---|---|
| Apple M1 Max 32GB | Mac | 32 GB (~21.4 usable) | ~18 tok/s | 8K · 31K (8-bit) | No price yet |
| Apple M1 Pro 32GB | Mac | 32 GB (~21.4 usable) | ~9 tok/s | 8K · 31K (8-bit) | No price yet |
| Apple M2 Max (30-core GPU) 32GB | Mac | 32 GB (~21.4 usable) | ~18 tok/s | 8K · 31K (8-bit) | No price yet |
| Apple M2 Max (38-core GPU) 32GB | Mac | 32 GB (~21.4 usable) | ~18 tok/s | 8K · 31K (8-bit) | No price yet |
| Apple M2 Pro 32GB | Mac | 32 GB (~21.4 usable) | ~9 tok/s | 8K · 31K (8-bit) | No price yet |
| Apple M4 32GB | Mac | 32 GB (~21.4 usable) | ~5 tok/s | 8K · 31K (8-bit) | No price yet |
| Apple M5 32GB | Mac | 32 GB (~21.4 usable) | ~7 tok/s | 8K · 31K (8-bit) | No price yet |
| Apple M6 32GB | Mac | 32 GB (~21.4 usable) | ~8 tok/s | 8K · 31K (8-bit) | No price yet |
| AMD Radeon RX 7900 XTX | GPU | 24 GB | ~38 tok/s | 32K (full) | low $1,350 |
| Intel Arc Pro B60 | Workstation GPU | 24 GB | ~14 tok/s | 32K (full) | No price yet |
| NVIDIA A10 | Data center GPU | 24 GB | ~26 tok/s | 32K (full) | No price yet |
| NVIDIA GeForce RTX 3090 | GPU | 24 GB | ~40 tok/s | 32K (full) | low $1,499 |
| NVIDIA GeForce RTX 3090 Ti | GPU | 24 GB | ~43 tok/s | 32K (full) | No price yet |
| NVIDIA GeForce RTX 4090 | GPU | 24 GB | ~43 tok/s | 32K (full) | low $2,910 |
| NVIDIA L4 | Data center GPU | 24 GB | ~13 tok/s | 32K (full) | No price yet |
| NVIDIA Quadro RTX 6000 | Workstation GPU | 24 GB | ~29 tok/s | 32K (full) | No price yet |
| NVIDIA RTX 4500 Ada Generation | Workstation GPU | 24 GB | ~19 tok/s | 32K (full) | No price yet |
| NVIDIA RTX A5000 | Workstation GPU | 24 GB | ~33 tok/s | 32K (full) | No price yet |
| NVIDIA RTX A5500 | Workstation GPU | 24 GB | ~33 tok/s | 32K (full) | No price yet |
| NVIDIA RTX PRO 4000 Blackwell | Workstation GPU | 24 GB | ~29 tok/s | 32K (full) | No price yet |
| NVIDIA RTX PRO 4000 SFF Blackwell | Workstation GPU | 24 GB | ~19 tok/s | 32K (full) | No price yet |
| NVIDIA RTX Spark N1X laptop (6144-core GPU, 24GB) | Laptop | 24 GB | ~13 tok/s | 32K (full) | No price yet |
| NVIDIA Tesla P40 | Data center GPU | 24 GB | ~11 tok/s | 32K (full) | No price yet |
| NVIDIA TITAN RTX | GPU | 24 GB | ~29 tok/s | 32K (full) | No price yet |
| Apple M3 Max (30-core GPU) 36GB | Mac | 36 GB (~24.1 usable) | ~13 tok/s | 32K (full) | No price yet |
| Apple M3 Pro 36GB | Mac | 36 GB (~24.1 usable) | ~7 tok/s | 32K (full) | No price yet |
| Apple M4 Max (32-core GPU) 36GB | Mac | 36 GB (~24.1 usable) | ~18 tok/s | 32K (full) | No price yet |
| Apple Mac Studio M5 Max (32-core GPU, 36 GB) | Mac | 36 GB (~24.1 usable) | ~20 tok/s | 32K (full) | low $2,499 |
| AMD Radeon PRO V710 | Data center GPU | 28 GB | ~18 tok/s | 32K (full) | No price yet |
| AMD Radeon AI PRO R9600 | Workstation GPU | 32 GB | ~25 tok/s | 32K (full) | No price yet |
| AMD Radeon AI PRO R9600D | Workstation GPU | 32 GB | ~25 tok/s | 32K (full) | No price yet |
| AMD Radeon AI PRO R9700 | Workstation GPU | 32 GB | ~25 tok/s | 32K (full) | No price yet |
| AMD Radeon AI PRO R9700S | Workstation GPU | 32 GB | ~25 tok/s | 32K (full) | No price yet |
| AMD Radeon PRO V620 | Data center GPU | 32 GB | ~20 tok/s | 32K (full) | No price yet |
| AMD Radeon PRO W6800 | Workstation GPU | 32 GB | ~20 tok/s | 32K (full) | No price yet |
| AMD Radeon PRO W7800 32GB | Workstation GPU | 32 GB | ~23 tok/s | 32K (full) | No price yet |
| AMD Ryzen AI Max 385 (32GB) | Mini PC | 32 GB | ~10 tok/s | 32K (full) | No price yet |
| AMD Ryzen AI Max 390 (32GB) | Mini PC | 32 GB | ~10 tok/s | 32K (full) | No price yet |
| AMD Ryzen AI Max+ 395 (32GB) | Mini PC | 32 GB | ~10 tok/s | 32K (full) | No price yet |
| Intel Arc Pro B70 | Workstation GPU | 32 GB | ~18 tok/s | 32K (full) | No price yet |
| NVIDIA GeForce RTX 5090 | GPU | 32 GB | ~74 tok/s | 32K (full) | low $6,500 |
| NVIDIA Quadro GV100 | Workstation GPU | 32 GB | ~37 tok/s | 32K (full) | No price yet |
| NVIDIA RTX 5000 Ada Generation | Workstation GPU | 32 GB | ~25 tok/s | 32K (full) | No price yet |
| NVIDIA RTX PRO 4500 Blackwell | Workstation GPU | 32 GB | ~39 tok/s | 32K (full) | No price yet |
| NVIDIA RTX Spark N1X laptop (6144-core GPU, 32GB) | Laptop | 32 GB | ~13 tok/s | 32K (full) | No price yet |
| NVIDIA V100 PCIe 32GB | Data center GPU | 32 GB | ~39 tok/s | 32K (full) | No price yet |
| NVIDIA V100S PCIe 32GB | Data center GPU | 32 GB | ~48 tok/s | 32K (full) | No price yet |
| Apple M3 Max (40-core GPU) 48GB | Mac | 48 GB (~36 usable) | ~18 tok/s | 32K (full) | No price yet |
| Apple M4 Max (40-core GPU) 48GB | Mac | 48 GB (~36 usable) | ~24 tok/s | 32K (full) | No price yet |
| Apple M4 Pro 48GB | Mac | 48 GB (~36 usable) | ~12 tok/s | 32K (full) | No price yet |
| Apple M5 Max (40-core GPU) 48GB | Mac | 48 GB (~36 usable) | ~27 tok/s | 32K (full) | No price yet |
| Apple M5 Pro 48GB | Mac | 48 GB (~36 usable) | ~14 tok/s | 32K (full) | No price yet |
| NVIDIA A100 40GB PCIe | Data center GPU | 40 GB | ~65 tok/s | 32K (full) | No price yet |
| NVIDIA A800 40GB Active | Workstation GPU | 40 GB | ~65 tok/s | 32K (full) | No price yet |
| AMD Radeon PRO W7800 48GB | Workstation GPU | 48 GB | ~34 tok/s | 32K (full) | No price yet |
| AMD Radeon PRO W7900 | Workstation GPU | 48 GB | ~34 tok/s | 32K (full) | No price yet |
| AMD Radeon PRO W7900 Dual Slot | Workstation GPU | 48 GB | ~34 tok/s | 32K (full) | No price yet |
| Apple M1 Max 64GB | Mac | 64 GB (~48 usable) | ~18 tok/s | 32K (full) | No price yet |
| Apple M1 Ultra 64GB | Mac | 64 GB (~48 usable) | ~34 tok/s | 32K (full) | No price yet |
| Apple M2 Max (30-core GPU) 64GB | Mac | 64 GB (~48 usable) | ~18 tok/s | 32K (full) | No price yet |
| Apple M2 Max (38-core GPU) 64GB | Mac | 64 GB (~48 usable) | ~18 tok/s | 32K (full) | No price yet |
| Apple M2 Ultra 64GB | Mac | 64 GB (~48 usable) | ~34 tok/s | 32K (full) | No price yet |
| Apple M3 Max (40-core GPU) 64GB | Mac | 64 GB (~48 usable) | ~18 tok/s | 32K (full) | No price yet |
| Apple M4 Max (40-core GPU) 64GB | Mac | 64 GB (~48 usable) | ~24 tok/s | 32K (full) | No price yet |
| Apple M4 Pro 64GB | Mac | 64 GB (~48 usable) | ~12 tok/s | 32K (full) | No price yet |
| Apple M5 Pro 64GB | Mac | 64 GB (~48 usable) | ~14 tok/s | 32K (full) | No price yet |
| Apple Mac Studio M5 Max (40-core GPU, 64 GB) | Mac | 64 GB (~48 usable) | ~27 tok/s | 32K (full) | No price yet |
| NVIDIA A40 | Data center GPU | 48 GB | ~30 tok/s | 32K (full) | No price yet |
| NVIDIA L40 | Data center GPU | 48 GB | ~37 tok/s | 32K (full) | No price yet |
| NVIDIA L40S | Data center GPU | 48 GB | ~37 tok/s | 32K (full) | No price yet |
| NVIDIA Quadro RTX 8000 | Workstation GPU | 48 GB | ~29 tok/s | 32K (full) | No price yet |
| NVIDIA RTX 6000 Ada Generation | Workstation GPU | 48 GB | ~41 tok/s | 32K (full) | No price yet |
| NVIDIA RTX A6000 | Workstation GPU | 48 GB | ~33 tok/s | 32K (full) | low $8,363 |
| NVIDIA RTX PRO 5000 Blackwell 48GB | Workstation GPU | 48 GB | ~56 tok/s | 32K (full) | No price yet |
| NVIDIA RTX Spark N1X laptop (6144-core GPU, 48GB) | Laptop | 48 GB | ~13 tok/s | 32K (full) | No price yet |
| AMD Ryzen AI Max 385 (64GB) | Mini PC | 64 GB | ~10 tok/s | 32K (full) | No price yet |
| AMD Ryzen AI Max 390 (64GB) | Mini PC | 64 GB | ~10 tok/s | 32K (full) | No price yet |
| AMD Ryzen AI Max+ 395 (64GB) | Mini PC | 64 GB | ~10 tok/s | 32K (full) | low $2,199.99 |
| NVIDIA DGX Spark (64GB) | Mini PC | 64 GB | ~12 tok/s | 32K (full) | low $4,999 |
| NVIDIA RTX Spark N1X laptop (6144-core GPU, 64GB) | Laptop | 64 GB | ~13 tok/s | 32K (full) | No price yet |
| Apple M2 Max (38-core GPU) 96GB | Mac | 96 GB (~72 usable) | ~18 tok/s | 32K (full) | No price yet |
| Apple M3 Max (30-core GPU) 96GB | Mac | 96 GB (~72 usable) | ~13 tok/s | 32K (full) | No price yet |
| Apple M3 Ultra 96GB | Mac | 96 GB (~72 usable) | ~35 tok/s | 32K (full) | No price yet |
| Apple Mac Studio M5 Ultra (64-core GPU, 96 GB) | Mac | 96 GB (~72 usable) | ~50 tok/s | 32K (full) | low $5,499 |
| NVIDIA RTX PRO 5000 Blackwell 72GB | Workstation GPU | 72 GB | ~56 tok/s | 32K (full) | No price yet |
| NVIDIA A100 80GB PCIe | Data center GPU | 80 GB | ~79 tok/s | 32K (full) | No price yet |
| Apple M1 Ultra 128GB | Mac | 128 GB (~96 usable) | ~34 tok/s | 32K (full) | No price yet |
| Apple M2 Ultra 128GB | Mac | 128 GB (~96 usable) | ~34 tok/s | 32K (full) | No price yet |
| Apple M3 Max (40-core GPU) 128GB | Mac | 128 GB (~96 usable) | ~18 tok/s | 32K (full) | No price yet |
| Apple M4 Max (40-core GPU) 128GB | Mac | 128 GB (~96 usable) | ~24 tok/s | 32K (full) | No price yet |
| Apple M5 Max (40-core GPU) 128GB | Mac | 128 GB (~96 usable) | ~27 tok/s | 32K (full) | No price yet |
| NVIDIA RTX PRO 6000 Blackwell Max-Q | Workstation GPU | 96 GB | ~74 tok/s | 32K (full) | No price yet |
| NVIDIA RTX PRO 6000 Blackwell Server Edition | Data center GPU | 96 GB | ~67 tok/s | 32K (full) | No price yet |
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition | Workstation GPU | 96 GB | ~74 tok/s | 32K (full) | No price yet |
| AMD Ryzen AI Max 385 (128GB) | Mini PC | 128 GB | ~10 tok/s | 32K (full) | No price yet |
| AMD Ryzen AI Max 390 (128GB) | Mini PC | 128 GB | ~10 tok/s | 32K (full) | No price yet |
| AMD Ryzen AI Max+ 395 (128GB) | Mini PC | 128 GB | ~10 tok/s | 32K (full) | low $3,499.99 |
| NVIDIA DGX Spark (128GB) | Mini PC | 128 GB | ~12 tok/s | 32K (full) | low $6,949.99 |
| NVIDIA RTX Spark N1X laptop (6144-core GPU, 128GB) | Laptop | 128 GB | ~13 tok/s | 32K (full) | No price yet |
| Apple M2 Ultra 192GB | Mac | 192 GB (~144 usable) | ~34 tok/s | 32K (full) | No price yet |
| Apple M3 Ultra 256GB | Mac | 256 GB (~192 usable) | ~35 tok/s | 32K (full) | No price yet |
| Apple M5 Ultra 256GB | Mac | 256 GB (~192 usable) | ~50 tok/s | 32K (full) | No price yet |
| Framework Desktop, AMD Ryzen AI Max+ PRO 495 (Strix Halo), 192 GB | Mini PC | 192 GB | ~11 tok/s | 32K (full) | low $6,799 |
| Apple M5 Ultra 512GB | Mac | 512 GB (~384 usable) | ~50 tok/s | 32K (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 | ~14 tok/s | 32K (full) |
| 2 × NVIDIA GeForce RTX 3060 12 GB | 24 GB | ~16 tok/s | 32K (full) |
| 2 × NVIDIA GeForce RTX 4070 SUPER | 24 GB | ~22 tok/s | 32K (full) |
| 3 × NVIDIA GeForce RTX 5060 | 24 GB | ~20 tok/s | 32K (full) |
| 3 × NVIDIA GeForce RTX 5060 Ti 8 GB | 24 GB | ~20 tok/s | 32K (full) |
| 2 × NVIDIA GeForce RTX 5070 | 24 GB | ~29 tok/s | 32K (full) |
| 2 × AMD Radeon RX 7800 XT | 32 GB | ~25 tok/s | 32K (full) |
| 2 × AMD Radeon RX 9060 XT 16 GB | 32 GB | ~13 tok/s | 32K (full) |
| 2 × AMD Radeon RX 9070 | 32 GB | ~25 tok/s | 32K (full) |
| 2 × AMD Radeon RX 9070 XT | 32 GB | ~25 tok/s | 32K (full) |
| 2 × NVIDIA GeForce RTX 5060 Ti 16 GB | 32 GB | ~20 tok/s | 32K (full) |
| 2 × NVIDIA GeForce RTX 5070 Ti | 32 GB | ~39 tok/s | 32K (full) |
| 2 × NVIDIA GeForce RTX 5080 | 32 GB | ~41 tok/s | 32K (full) |
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
Related
- BuildBuild that fits it: 24 GB: one used RTX 3090
- BuildBuild that fits it: 32 GB: two Radeon RX 9070
- BuildBuild that fits it: 48 GB: dual used RTX 3090
- ModelPLaMo 3 Translate 0.8B, another size of the same translation model
- ModelPLaMo 3 Translate 2B, another size of the same translation model
- ModelPLaMo 3 Translate 8B, another size of the same translation model
- CatalogEvery model in the catalog