Model field guide
Qwen3-4B-Instruct-2507
42 of 44 hardware configurations have a supported memory fit at one or more recorded quantizations. Runtime compatibility still needs verification.
Total parameters4.02B
Active parameters4.02B
Recorded context limit262,144
Architecturedense
Verified 2026-10-22 · license metadata: apache-2.0 · commercial use: yes (review the terms).
Actual files, not a size shortcut.
| Quantization | Total file size | Source |
|---|---|---|
| Q4_K_M | 2.326 GiB2,497,281,120 bytes; 1 file(s)Sum of exact file byte sizes in the repository listing at the recorded commit; every declared shard is present. Auxiliary vision projector and speculative-decoding files are excluded. Labels follow publisher filenames; mixed-precision expert tensors may remain in a differently labeled file, so the estimator uses measured bytes rather than assumed bits per weight. | Pinned file recordVerified 2026-10-22 |
| Q5_K_M | 2.691 GiB2,889,514,080 bytes; 1 file(s)Sum of exact file byte sizes in the repository listing at the recorded commit; every declared shard is present. Auxiliary vision projector and speculative-decoding files are excluded. Labels follow publisher filenames; mixed-precision expert tensors may remain in a differently labeled file, so the estimator uses measured bytes rather than assumed bits per weight. | Pinned file recordVerified 2026-10-22 |
| Q8_0 | 3.986 GiB4,280,405,600 bytes; 1 file(s)Sum of exact file byte sizes in the repository listing at the recorded commit; every declared shard is present. Auxiliary vision projector and speculative-decoding files are excluded. Labels follow publisher filenames; mixed-precision expert tensors may remain in a differently labeled file, so the estimator uses measured bytes rather than assumed bits per weight. | Pinned file recordVerified 2026-10-22 |
| FP16 | 7.498 GiB8,051,285,344 bytes; 1 file(s)Sum of exact file byte sizes in the repository listing at the recorded commit; every declared shard is present. Auxiliary vision projector and speculative-decoding files are excluded. Labels follow publisher filenames; mixed-precision expert tensors may remain in a differently labeled file, so the estimator uses measured bytes rather than assumed bits per weight. | Pinned file recordVerified 2026-10-22 |
The hardware picture.
Q4_K_M: hardware comparison
| Hardware | Usable memory | Required memory | Fit estimate | Estimated generation |
|---|---|---|---|---|
| Mac mini M6 (12-core GPU) 16GB | 12.0 GiB | 3.89 GiB | fits | 24.5–49.0 estimated tokens/s |
| Mac mini M6 (12-core GPU) 24GB | 18.0 GiB | 3.89 GiB | fits | 27.2–54.5 estimated tokens/s |
| Mac mini M6 (12-core GPU) 32GB | 24.0 GiB | 3.89 GiB | fits | 27.2–54.5 estimated tokens/s |
| Mac mini M5 Pro (15-core CPU / 16-core GPU) 24GB | 18.0 GiB | 3.89 GiB | fits | 49.2–98.3 estimated tokens/s |
| Mac mini M5 Pro (15-core CPU / 16-core GPU) 48GB | 36.0 GiB | 3.89 GiB | fits | 49.2–98.3 estimated tokens/s |
| Mac mini M5 Pro (15-core CPU / 16-core GPU) 64GB | 48.0 GiB | 3.89 GiB | fits | 49.2–98.3 estimated tokens/s |
| Mac Studio M5 Max (32-core GPU) 36GB | 27.0 GiB | 3.89 GiB | fits | 73.7–147.4 estimated tokens/s |
| Mac Studio M5 Max (40-core GPU) 64GB | 48.0 GiB | 3.89 GiB | fits | 98.3–196.7 estimated tokens/s |
| Mac Studio M5 Max (40-core GPU) 128GB | 96.0 GiB | 3.89 GiB | fits | 98.3–196.7 estimated tokens/s |
| Mac Studio M5 Ultra (30-core CPU / 64-core GPU) 96GB | 72.0 GiB | 3.89 GiB | fits | 192.2–384.4 estimated tokens/s |
| Mac Studio M5 Ultra (36-core CPU / 80-core GPU) 256GB | 192.0 GiB | 3.89 GiB | fits | 192.2–384.4 estimated tokens/s |
| Mac Studio M5 Ultra (36-core CPU / 80-core GPU) 512GB | 384.0 GiB | 3.89 GiB | fits | 192.2–384.4 estimated tokens/s |
| MacBook Pro 14-inch M5 (10-core GPU) 24GB | 18.0 GiB | 3.89 GiB | fits | 24.5–49.0 estimated tokens/s |
| MacBook Pro 14-inch M5 Pro (20-core GPU) 48GB | 36.0 GiB | 3.89 GiB | fits | 49.2–98.3 estimated tokens/s |
| MacBook Pro 16-inch M5 Max (40-core GPU) 128GB | 96.0 GiB | 3.89 GiB | fits | 98.3–196.7 estimated tokens/s |
| Mac mini (2024) M4 (10-core GPU) 16GB | 12.0 GiB | 3.89 GiB | fits | 19.2–38.4 estimated tokens/s |
| Mac mini (2024) M4 (10-core GPU) 24GB | 18.0 GiB | 3.89 GiB | fits | 19.2–38.4 estimated tokens/s |
| Mac mini (2024) M4 (10-core GPU) 32GB | 24.0 GiB | 3.89 GiB | fits | 19.2–38.4 estimated tokens/s |
| Mac mini (2024) M4 Pro (16-core GPU) 24GB | 18.0 GiB | 3.89 GiB | fits | 43.7–87.5 estimated tokens/s |
| Mac mini (2024) M4 Pro (16-core GPU) 48GB | 36.0 GiB | 3.89 GiB | fits | 43.7–87.5 estimated tokens/s |
| Mac mini (2024) M4 Pro (16-core GPU) 64GB | 48.0 GiB | 3.89 GiB | fits | 43.7–87.5 estimated tokens/s |
| Mac Studio (2025) M4 Max (32-core GPU) 36GB | 27.0 GiB | 3.89 GiB | fits | 65.7–131.3 estimated tokens/s |
| Mac Studio (2025) M4 Max (40-core GPU) 64GB | 48.0 GiB | 3.89 GiB | fits | 87.5–174.9 estimated tokens/s |
| Mac Studio (2025) M4 Max (40-core GPU) 128GB | 96.0 GiB | 3.89 GiB | fits | 87.5–174.9 estimated tokens/s |
| Mac Studio (2025) M3 Ultra (28-core CPU / 60-core GPU) 96GB | 72.0 GiB | 3.89 GiB | fits | 131.2–262.4 estimated tokens/s |
| Mac Studio (2025) M3 Ultra (32-core CPU / 80-core GPU) 256GB | 192.0 GiB | 3.89 GiB | fits | 131.2–262.4 estimated tokens/s |
| NVIDIA GeForce RTX 3060 12GB | 10.8 GiB | 3.89 GiB | fits | 57.7–115.3 estimated tokens/s |
| NVIDIA GeForce RTX 3090 24GB | 21.6 GiB | 3.89 GiB | fits | 149.9–299.8 estimated tokens/s |
| NVIDIA GeForce RTX 4060 Ti 16GB | 14.4 GiB | 3.89 GiB | fits | 46.1–92.3 estimated tokens/s |
| NVIDIA GeForce RTX 4070 12GB GDDR6X | 10.8 GiB | 3.89 GiB | fits | 80.7–161.5 estimated tokens/s |
| NVIDIA GeForce RTX 4080 16GB | 14.4 GiB | 3.89 GiB | fits | 114.8–229.6 estimated tokens/s |
| NVIDIA GeForce RTX 4090 24GB | 21.6 GiB | 3.89 GiB | fits | 161.5–322.9 estimated tokens/s |
| NVIDIA GeForce RTX 5060 Ti 16GB | 14.4 GiB | 3.89 GiB | fits | 71.8–143.5 estimated tokens/s |
| NVIDIA GeForce RTX 5070 12GB | 10.8 GiB | 3.89 GiB | fits | 107.6–215.3 estimated tokens/s |
| NVIDIA GeForce RTX 5070 Ti 16GB | 14.4 GiB | 3.89 GiB | fits | 143.5–287.0 estimated tokens/s |
| NVIDIA GeForce RTX 5080 16GB | 14.4 GiB | 3.89 GiB | fits | 153.8–307.5 estimated tokens/s |
| NVIDIA GeForce RTX 5090 32GB | 28.8 GiB | 3.89 GiB | fits | 287.0–574.1 estimated tokens/s |
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition 96GB | 86.4 GiB | 3.89 GiB | fits | 287.0–574.1 estimated tokens/s |
| NVIDIA RTX 6000 Ada Generation 48GB | 43.2 GiB | 3.89 GiB | fits | 153.8–307.5 estimated tokens/s |
| Intel Arc B580 12GB | 10.8 GiB | 3.89 GiB | fits | 73.0–146.1 estimated tokens/s |
| AMD Ryzen AI Halo, Ryzen AI Max+ 395, 128GB | 89.6 GiB | 3.89 GiB | fits | 41.0–82.0 estimated tokens/s |
| NVIDIA DGX Spark 128GB / 4TB | 96.0 GiB | 3.89 GiB | fits | 43.7–87.5 estimated tokens/s |
| 2 × GeForce RTX 3090 24GB (conditional 48GB aggregate) | 40.8 GiB | 3.89 GiB | Needs verification | Estimated speed unavailable |
| 2 × GeForce RTX 5090 32GB (conditional 64GB aggregate) | 54.4 GiB | 3.89 GiB | Needs verification | Estimated speed unavailable |
Q5_K_M: hardware comparison
| Hardware | Usable memory | Required memory | Fit estimate | Estimated generation |
|---|---|---|---|---|
| Mac mini M6 (12-core GPU) 16GB | 12.0 GiB | 4.25 GiB | fits | 21.2–42.4 estimated tokens/s |
| Mac mini M6 (12-core GPU) 24GB | 18.0 GiB | 4.25 GiB | fits | 23.5–47.1 estimated tokens/s |
| Mac mini M6 (12-core GPU) 32GB | 24.0 GiB | 4.25 GiB | fits | 23.5–47.1 estimated tokens/s |
| Mac mini M5 Pro (15-core CPU / 16-core GPU) 24GB | 18.0 GiB | 4.25 GiB | fits | 42.5–85.0 estimated tokens/s |
| Mac mini M5 Pro (15-core CPU / 16-core GPU) 48GB | 36.0 GiB | 4.25 GiB | fits | 42.5–85.0 estimated tokens/s |
| Mac mini M5 Pro (15-core CPU / 16-core GPU) 64GB | 48.0 GiB | 4.25 GiB | fits | 42.5–85.0 estimated tokens/s |
| Mac Studio M5 Max (32-core GPU) 36GB | 27.0 GiB | 4.25 GiB | fits | 63.7–127.4 estimated tokens/s |
| Mac Studio M5 Max (40-core GPU) 64GB | 48.0 GiB | 4.25 GiB | fits | 85.0–170.0 estimated tokens/s |
| Mac Studio M5 Max (40-core GPU) 128GB | 96.0 GiB | 4.25 GiB | fits | 85.0–170.0 estimated tokens/s |
| Mac Studio M5 Ultra (30-core CPU / 64-core GPU) 96GB | 72.0 GiB | 4.25 GiB | fits | 166.1–332.2 estimated tokens/s |
| Mac Studio M5 Ultra (36-core CPU / 80-core GPU) 256GB | 192.0 GiB | 4.25 GiB | fits | 166.1–332.2 estimated tokens/s |
| Mac Studio M5 Ultra (36-core CPU / 80-core GPU) 512GB | 384.0 GiB | 4.25 GiB | fits | 166.1–332.2 estimated tokens/s |
| MacBook Pro 14-inch M5 (10-core GPU) 24GB | 18.0 GiB | 4.25 GiB | fits | 21.2–42.4 estimated tokens/s |
| MacBook Pro 14-inch M5 Pro (20-core GPU) 48GB | 36.0 GiB | 4.25 GiB | fits | 42.5–85.0 estimated tokens/s |
| MacBook Pro 16-inch M5 Max (40-core GPU) 128GB | 96.0 GiB | 4.25 GiB | fits | 85.0–170.0 estimated tokens/s |
| Mac mini (2024) M4 (10-core GPU) 16GB | 12.0 GiB | 4.25 GiB | fits | 16.6–33.2 estimated tokens/s |
| Mac mini (2024) M4 (10-core GPU) 24GB | 18.0 GiB | 4.25 GiB | fits | 16.6–33.2 estimated tokens/s |
| Mac mini (2024) M4 (10-core GPU) 32GB | 24.0 GiB | 4.25 GiB | fits | 16.6–33.2 estimated tokens/s |
| Mac mini (2024) M4 Pro (16-core GPU) 24GB | 18.0 GiB | 4.25 GiB | fits | 37.8–75.6 estimated tokens/s |
| Mac mini (2024) M4 Pro (16-core GPU) 48GB | 36.0 GiB | 4.25 GiB | fits | 37.8–75.6 estimated tokens/s |
| Mac mini (2024) M4 Pro (16-core GPU) 64GB | 48.0 GiB | 4.25 GiB | fits | 37.8–75.6 estimated tokens/s |
| Mac Studio (2025) M4 Max (32-core GPU) 36GB | 27.0 GiB | 4.25 GiB | fits | 56.8–113.5 estimated tokens/s |
| Mac Studio (2025) M4 Max (40-core GPU) 64GB | 48.0 GiB | 4.25 GiB | fits | 75.6–151.2 estimated tokens/s |
| Mac Studio (2025) M4 Max (40-core GPU) 128GB | 96.0 GiB | 4.25 GiB | fits | 75.6–151.2 estimated tokens/s |
| Mac Studio (2025) M3 Ultra (28-core CPU / 60-core GPU) 96GB | 72.0 GiB | 4.25 GiB | fits | 113.4–226.8 estimated tokens/s |
| Mac Studio (2025) M3 Ultra (32-core CPU / 80-core GPU) 256GB | 192.0 GiB | 4.25 GiB | fits | 113.4–226.8 estimated tokens/s |
| NVIDIA GeForce RTX 3060 12GB | 10.8 GiB | 4.25 GiB | fits | 49.8–99.7 estimated tokens/s |
| NVIDIA GeForce RTX 3090 24GB | 21.6 GiB | 4.25 GiB | fits | 129.6–259.1 estimated tokens/s |
| NVIDIA GeForce RTX 4060 Ti 16GB | 14.4 GiB | 4.25 GiB | fits | 39.9–79.7 estimated tokens/s |
| NVIDIA GeForce RTX 4070 12GB GDDR6X | 10.8 GiB | 4.25 GiB | fits | 69.8–139.5 estimated tokens/s |
| NVIDIA GeForce RTX 4080 16GB | 14.4 GiB | 4.25 GiB | fits | 99.2–198.5 estimated tokens/s |
| NVIDIA GeForce RTX 4090 24GB | 21.6 GiB | 4.25 GiB | fits | 139.5–279.1 estimated tokens/s |
| NVIDIA GeForce RTX 5060 Ti 16GB | 14.4 GiB | 4.25 GiB | fits | 62.0–124.0 estimated tokens/s |
| NVIDIA GeForce RTX 5070 12GB | 10.8 GiB | 4.25 GiB | fits | 93.0–186.1 estimated tokens/s |
| NVIDIA GeForce RTX 5070 Ti 16GB | 14.4 GiB | 4.25 GiB | fits | 124.0–248.1 estimated tokens/s |
| NVIDIA GeForce RTX 5080 16GB | 14.4 GiB | 4.25 GiB | fits | 132.9–265.8 estimated tokens/s |
| NVIDIA GeForce RTX 5090 32GB | 28.8 GiB | 4.25 GiB | fits | 248.1–496.1 estimated tokens/s |
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition 96GB | 86.4 GiB | 4.25 GiB | fits | 248.1–496.1 estimated tokens/s |
| NVIDIA RTX 6000 Ada Generation 48GB | 43.2 GiB | 4.25 GiB | fits | 132.9–265.8 estimated tokens/s |
| Intel Arc B580 12GB | 10.8 GiB | 4.25 GiB | fits | 63.1–126.2 estimated tokens/s |
| AMD Ryzen AI Halo, Ryzen AI Max+ 395, 128GB | 89.6 GiB | 4.25 GiB | fits | 35.4–70.9 estimated tokens/s |
| NVIDIA DGX Spark 128GB / 4TB | 96.0 GiB | 4.25 GiB | fits | 37.8–75.6 estimated tokens/s |
| 2 × GeForce RTX 3090 24GB (conditional 48GB aggregate) | 40.8 GiB | 4.25 GiB | Needs verification | Estimated speed unavailable |
| 2 × GeForce RTX 5090 32GB (conditional 64GB aggregate) | 54.4 GiB | 4.25 GiB | Needs verification | Estimated speed unavailable |
Q8_0: hardware comparison
| Hardware | Usable memory | Required memory | Fit estimate | Estimated generation |
|---|---|---|---|---|
| Mac mini M6 (12-core GPU) 16GB | 12.0 GiB | 5.55 GiB | fits | 14.3–28.6 estimated tokens/s |
| Mac mini M6 (12-core GPU) 24GB | 18.0 GiB | 5.55 GiB | fits | 15.9–31.8 estimated tokens/s |
| Mac mini M6 (12-core GPU) 32GB | 24.0 GiB | 5.55 GiB | fits | 15.9–31.8 estimated tokens/s |
| Mac mini M5 Pro (15-core CPU / 16-core GPU) 24GB | 18.0 GiB | 5.55 GiB | fits | 28.7–57.4 estimated tokens/s |
| Mac mini M5 Pro (15-core CPU / 16-core GPU) 48GB | 36.0 GiB | 5.55 GiB | fits | 28.7–57.4 estimated tokens/s |
| Mac mini M5 Pro (15-core CPU / 16-core GPU) 64GB | 48.0 GiB | 5.55 GiB | fits | 28.7–57.4 estimated tokens/s |
| Mac Studio M5 Max (32-core GPU) 36GB | 27.0 GiB | 5.55 GiB | fits | 43.0–86.0 estimated tokens/s |
| Mac Studio M5 Max (40-core GPU) 64GB | 48.0 GiB | 5.55 GiB | fits | 57.4–114.8 estimated tokens/s |
| Mac Studio M5 Max (40-core GPU) 128GB | 96.0 GiB | 5.55 GiB | fits | 57.4–114.8 estimated tokens/s |
| Mac Studio M5 Ultra (30-core CPU / 64-core GPU) 96GB | 72.0 GiB | 5.55 GiB | fits | 112.1–224.3 estimated tokens/s |
| Mac Studio M5 Ultra (36-core CPU / 80-core GPU) 256GB | 192.0 GiB | 5.55 GiB | fits | 112.1–224.3 estimated tokens/s |
| Mac Studio M5 Ultra (36-core CPU / 80-core GPU) 512GB | 384.0 GiB | 5.55 GiB | fits | 112.1–224.3 estimated tokens/s |
| MacBook Pro 14-inch M5 (10-core GPU) 24GB | 18.0 GiB | 5.55 GiB | fits | 14.3–28.6 estimated tokens/s |
| MacBook Pro 14-inch M5 Pro (20-core GPU) 48GB | 36.0 GiB | 5.55 GiB | fits | 28.7–57.4 estimated tokens/s |
| MacBook Pro 16-inch M5 Max (40-core GPU) 128GB | 96.0 GiB | 5.55 GiB | fits | 57.4–114.8 estimated tokens/s |
| Mac mini (2024) M4 (10-core GPU) 16GB | 12.0 GiB | 5.55 GiB | fits | 11.2–22.4 estimated tokens/s |
| Mac mini (2024) M4 (10-core GPU) 24GB | 18.0 GiB | 5.55 GiB | fits | 11.2–22.4 estimated tokens/s |
| Mac mini (2024) M4 (10-core GPU) 32GB | 24.0 GiB | 5.55 GiB | fits | 11.2–22.4 estimated tokens/s |
| Mac mini (2024) M4 Pro (16-core GPU) 24GB | 18.0 GiB | 5.55 GiB | fits | 25.5–51.0 estimated tokens/s |
| Mac mini (2024) M4 Pro (16-core GPU) 48GB | 36.0 GiB | 5.55 GiB | fits | 25.5–51.0 estimated tokens/s |
| Mac mini (2024) M4 Pro (16-core GPU) 64GB | 48.0 GiB | 5.55 GiB | fits | 25.5–51.0 estimated tokens/s |
| Mac Studio (2025) M4 Max (32-core GPU) 36GB | 27.0 GiB | 5.55 GiB | fits | 38.3–76.6 estimated tokens/s |
| Mac Studio (2025) M4 Max (40-core GPU) 64GB | 48.0 GiB | 5.55 GiB | fits | 51.0–102.0 estimated tokens/s |
| Mac Studio (2025) M4 Max (40-core GPU) 128GB | 96.0 GiB | 5.55 GiB | fits | 51.0–102.0 estimated tokens/s |
| Mac Studio (2025) M3 Ultra (28-core CPU / 60-core GPU) 96GB | 72.0 GiB | 5.55 GiB | fits | 76.5–153.1 estimated tokens/s |
| Mac Studio (2025) M3 Ultra (32-core CPU / 80-core GPU) 256GB | 192.0 GiB | 5.55 GiB | fits | 76.5–153.1 estimated tokens/s |
| NVIDIA GeForce RTX 3060 12GB | 10.8 GiB | 5.55 GiB | fits | 33.6–67.3 estimated tokens/s |
| NVIDIA GeForce RTX 3090 24GB | 21.6 GiB | 5.55 GiB | fits | 87.5–174.9 estimated tokens/s |
| NVIDIA GeForce RTX 4060 Ti 16GB | 14.4 GiB | 5.55 GiB | fits | 26.9–53.8 estimated tokens/s |
| NVIDIA GeForce RTX 4070 12GB GDDR6X | 10.8 GiB | 5.55 GiB | fits | 47.1–94.2 estimated tokens/s |
| NVIDIA GeForce RTX 4080 16GB | 14.4 GiB | 5.55 GiB | fits | 67.0–134.0 estimated tokens/s |
| NVIDIA GeForce RTX 4090 24GB | 21.6 GiB | 5.55 GiB | fits | 94.2–188.4 estimated tokens/s |
| NVIDIA GeForce RTX 5060 Ti 16GB | 14.4 GiB | 5.55 GiB | fits | 41.9–83.7 estimated tokens/s |
| NVIDIA GeForce RTX 5070 12GB | 10.8 GiB | 5.55 GiB | fits | 62.8–125.6 estimated tokens/s |
| NVIDIA GeForce RTX 5070 Ti 16GB | 14.4 GiB | 5.55 GiB | fits | 83.7–167.5 estimated tokens/s |
| NVIDIA GeForce RTX 5080 16GB | 14.4 GiB | 5.55 GiB | fits | 89.7–179.4 estimated tokens/s |
| NVIDIA GeForce RTX 5090 32GB | 28.8 GiB | 5.55 GiB | fits | 167.5–334.9 estimated tokens/s |
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition 96GB | 86.4 GiB | 5.55 GiB | fits | 167.5–334.9 estimated tokens/s |
| NVIDIA RTX 6000 Ada Generation 48GB | 43.2 GiB | 5.55 GiB | fits | 89.7–179.4 estimated tokens/s |
| Intel Arc B580 12GB | 10.8 GiB | 5.55 GiB | fits | 42.6–85.2 estimated tokens/s |
| AMD Ryzen AI Halo, Ryzen AI Max+ 395, 128GB | 89.6 GiB | 5.55 GiB | fits | 23.9–47.8 estimated tokens/s |
| NVIDIA DGX Spark 128GB / 4TB | 96.0 GiB | 5.55 GiB | fits | 25.5–51.0 estimated tokens/s |
| 2 × GeForce RTX 3090 24GB (conditional 48GB aggregate) | 40.8 GiB | 5.55 GiB | Needs verification | Estimated speed unavailable |
| 2 × GeForce RTX 5090 32GB (conditional 64GB aggregate) | 54.4 GiB | 5.55 GiB | Needs verification | Estimated speed unavailable |
FP16: hardware comparison
| Hardware | Usable memory | Required memory | Fit estimate | Estimated generation |
|---|---|---|---|---|
| Mac mini M6 (12-core GPU) 16GB | 12.0 GiB | 9.06 GiB | fits | 7.6–15.2 estimated tokens/s |
| Mac mini M6 (12-core GPU) 24GB | 18.0 GiB | 9.06 GiB | fits | 8.4–16.9 estimated tokens/s |
| Mac mini M6 (12-core GPU) 32GB | 24.0 GiB | 9.06 GiB | fits | 8.4–16.9 estimated tokens/s |
| Mac mini M5 Pro (15-core CPU / 16-core GPU) 24GB | 18.0 GiB | 9.06 GiB | fits | 15.3–30.5 estimated tokens/s |
| Mac mini M5 Pro (15-core CPU / 16-core GPU) 48GB | 36.0 GiB | 9.06 GiB | fits | 15.3–30.5 estimated tokens/s |
| Mac mini M5 Pro (15-core CPU / 16-core GPU) 64GB | 48.0 GiB | 9.06 GiB | fits | 15.3–30.5 estimated tokens/s |
| Mac Studio M5 Max (32-core GPU) 36GB | 27.0 GiB | 9.06 GiB | fits | 22.9–45.7 estimated tokens/s |
| Mac Studio M5 Max (40-core GPU) 64GB | 48.0 GiB | 9.06 GiB | fits | 30.5–61.0 estimated tokens/s |
| Mac Studio M5 Max (40-core GPU) 128GB | 96.0 GiB | 9.06 GiB | fits | 30.5–61.0 estimated tokens/s |
| Mac Studio M5 Ultra (30-core CPU / 64-core GPU) 96GB | 72.0 GiB | 9.06 GiB | fits | 59.6–119.2 estimated tokens/s |
| Mac Studio M5 Ultra (36-core CPU / 80-core GPU) 256GB | 192.0 GiB | 9.06 GiB | fits | 59.6–119.2 estimated tokens/s |
| Mac Studio M5 Ultra (36-core CPU / 80-core GPU) 512GB | 384.0 GiB | 9.06 GiB | fits | 59.6–119.2 estimated tokens/s |
| MacBook Pro 14-inch M5 (10-core GPU) 24GB | 18.0 GiB | 9.06 GiB | fits | 7.6–15.2 estimated tokens/s |
| MacBook Pro 14-inch M5 Pro (20-core GPU) 48GB | 36.0 GiB | 9.06 GiB | fits | 15.3–30.5 estimated tokens/s |
| MacBook Pro 16-inch M5 Max (40-core GPU) 128GB | 96.0 GiB | 9.06 GiB | fits | 30.5–61.0 estimated tokens/s |
| Mac mini (2024) M4 (10-core GPU) 16GB | 12.0 GiB | 9.06 GiB | fits | 6.0–11.9 estimated tokens/s |
| Mac mini (2024) M4 (10-core GPU) 24GB | 18.0 GiB | 9.06 GiB | fits | 6.0–11.9 estimated tokens/s |
| Mac mini (2024) M4 (10-core GPU) 32GB | 24.0 GiB | 9.06 GiB | fits | 6.0–11.9 estimated tokens/s |
| Mac mini (2024) M4 Pro (16-core GPU) 24GB | 18.0 GiB | 9.06 GiB | fits | 13.6–27.1 estimated tokens/s |
| Mac mini (2024) M4 Pro (16-core GPU) 48GB | 36.0 GiB | 9.06 GiB | fits | 13.6–27.1 estimated tokens/s |
| Mac mini (2024) M4 Pro (16-core GPU) 64GB | 48.0 GiB | 9.06 GiB | fits | 13.6–27.1 estimated tokens/s |
| Mac Studio (2025) M4 Max (32-core GPU) 36GB | 27.0 GiB | 9.06 GiB | fits | 20.4–40.7 estimated tokens/s |
| Mac Studio (2025) M4 Max (40-core GPU) 64GB | 48.0 GiB | 9.06 GiB | fits | 27.1–54.3 estimated tokens/s |
| Mac Studio (2025) M4 Max (40-core GPU) 128GB | 96.0 GiB | 9.06 GiB | fits | 27.1–54.3 estimated tokens/s |
| Mac Studio (2025) M3 Ultra (28-core CPU / 60-core GPU) 96GB | 72.0 GiB | 9.06 GiB | fits | 40.7–81.4 estimated tokens/s |
| Mac Studio (2025) M3 Ultra (32-core CPU / 80-core GPU) 256GB | 192.0 GiB | 9.06 GiB | fits | 40.7–81.4 estimated tokens/s |
| NVIDIA GeForce RTX 3060 12GB | 10.8 GiB | 9.06 GiB | fits | 17.9–35.8 estimated tokens/s |
| NVIDIA GeForce RTX 3090 24GB | 21.6 GiB | 9.06 GiB | fits | 46.5–93.0 estimated tokens/s |
| NVIDIA GeForce RTX 4060 Ti 16GB | 14.4 GiB | 9.06 GiB | fits | 14.3–28.6 estimated tokens/s |
| NVIDIA GeForce RTX 4070 12GB GDDR6X | 10.8 GiB | 9.06 GiB | fits | 25.0–50.1 estimated tokens/s |
| NVIDIA GeForce RTX 4080 16GB | 14.4 GiB | 9.06 GiB | fits | 35.6–71.2 estimated tokens/s |
| NVIDIA GeForce RTX 4090 24GB | 21.6 GiB | 9.06 GiB | fits | 50.1–100.2 estimated tokens/s |
| NVIDIA GeForce RTX 5060 Ti 16GB | 14.4 GiB | 9.06 GiB | fits | 22.3–44.5 estimated tokens/s |
| NVIDIA GeForce RTX 5070 12GB | 10.8 GiB | 9.06 GiB | fits | 33.4–66.8 estimated tokens/s |
| NVIDIA GeForce RTX 5070 Ti 16GB | 14.4 GiB | 9.06 GiB | fits | 44.5–89.0 estimated tokens/s |
| NVIDIA GeForce RTX 5080 16GB | 14.4 GiB | 9.06 GiB | fits | 47.7–95.4 estimated tokens/s |
| NVIDIA GeForce RTX 5090 32GB | 28.8 GiB | 9.06 GiB | fits | 89.0–178.1 estimated tokens/s |
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition 96GB | 86.4 GiB | 9.06 GiB | fits | 89.0–178.1 estimated tokens/s |
| NVIDIA RTX 6000 Ada Generation 48GB | 43.2 GiB | 9.06 GiB | fits | 47.7–95.4 estimated tokens/s |
| Intel Arc B580 12GB | 10.8 GiB | 9.06 GiB | fits | 22.7–45.3 estimated tokens/s |
| AMD Ryzen AI Halo, Ryzen AI Max+ 395, 128GB | 89.6 GiB | 9.06 GiB | fits | 12.7–25.4 estimated tokens/s |
| NVIDIA DGX Spark 128GB / 4TB | 96.0 GiB | 9.06 GiB | fits | 13.6–27.1 estimated tokens/s |
| 2 × GeForce RTX 3090 24GB (conditional 48GB aggregate) | 40.8 GiB | 9.06 GiB | Needs verification | Estimated speed unavailable |
| 2 × GeForce RTX 5090 32GB (conditional 64GB aggregate) | 54.4 GiB | 9.06 GiB | Needs verification | Estimated speed unavailable |
A “fits” result checks estimated memory only. “Needs verification” preserves missing cache, vision, or placement information. All speed figures are estimated from the published bandwidth formula.
Field-level sources and verification dates
- id: Official source · verified 2026-10-22. Lowercase URL slug derived from the upstream repository ID; baseModel retains the exact Hugging Face identifier.
- name: Official source · verified 2026-10-22
- baseModel: Official source · verified 2026-10-22
- downloads: Official source · verified 2026-10-22
- tasks: Official source · verified 2026-10-22. Chat denotes a text-generating model, including base models that may require a chat template or instruction tuning. Code label is inferred from the repository model-family name. Vision uses pipeline metadata or vision_config.
- license: Official source · verified 2026-10-22
- commercialUse: Official source · verified 2026-10-22. Official upstream organization metadata identifies a permissive standard license. Commercial use is permitted subject to the license terms, notices, and applicable restrictions; this is a metadata-based classification, not an individualized legal approval.
- license_url: Official source · verified 2026-10-22. Pinned license file confirmed in repository listing.
- architecture: Official source · verified 2026-10-22. Classification from model_type and expert/state-space configuration fields (qwen3); hybrid classification takes precedence over expert count.
- parametersB: Official source · verified 2026-10-22. Exact safetensors parameter total reported by the Hugging Face API, divided by 1e9.
- activeParametersB: Official source · verified 2026-10-22. Dense model: all parameters are included. This is not a runtime FLOP measurement.
- contextLength: Official source · verified 2026-10-22. Configured maximum token positions; not a guarantee of runtime support or quality at that length. No speculative RoPE scaling applied.
- kvCache: Official source · verified 2026-10-22. Standard attention KV shape from num_hidden_layers, num_key_value_heads, and head_dim or hidden_size / num_attention_heads; FP16 cache bytes are a separate estimator choice.
- licenseStandard: Official source · verified 2026-10-22. Official standard license text supporting the permissive-license classification; model applicability is identified separately by upstream metadata.