Qwen3.5 4B
Small text and image model with reasoning and tool use.
Stored parameters 4.7B
Native context 256Ktokens
ArchitectureDense
Extended context: 1,010,000 tokens. Requires the publisher-documented YaRN/RoPE scaling configuration; selecting a longer context alone does not enable it.
Memory by precision and context
GiB · one request · FP16 cache · default single-device buffers. Extended contexts are marked *. Text inference only.
| Weight precision | Weights | 8K total | 32K total | 128K total | 256K total | 512K* total | 1,010,000* total |
|---|---|---|---|---|---|---|---|
| Q4 · smaller | 2.6 | 3.9 | 4.7 | 7.7 | 11.7 | 19.7 | 34.5 |
| Q5 · medium | 3.1 | 4.4 | 5.1 | 8.1 | 12.1 | 20.1 | 35.0 |
| Q8 · larger | 4.6 | 5.9 | 6.7 | 9.7 | 13.7 | 21.7 | 36.5 |
| FP16 / BF16 | 8.7 | 10.0 | 10.7 | 13.7 | 17.7 | 25.7 | 40.6 |
Hardware fit at Q4 and 8K
One request, FP16 cache. Multi-GPU and compressed or sliding caches require compatible runtime support.
View 199 hardware configurations
| Hardware | Memory needed | Available | Estimate | Open plan |
|---|---|---|---|---|
| GeForce RTX 50508 GiB · GPU | 3.9 GiB | 7.0 GiB | Room to spare | |
| GeForce RTX 50608 GiB · GPU | 3.9 GiB | 7.0 GiB | Room to spare | |
| GeForce RTX 5060 Ti 8 GB8 GiB · GPU | 3.9 GiB | 7.0 GiB | Room to spare | |
| GeForce RTX 5060 Ti 16 GB16 GiB · GPU | 3.9 GiB | 15.0 GiB | Room to spare | |
| GeForce RTX 507012 GiB · GPU | 3.9 GiB | 11.0 GiB | Room to spare | |
| GeForce RTX 5070 Ti16 GiB · GPU | 3.9 GiB | 15.0 GiB | Room to spare | |
| GeForce RTX 508016 GiB · GPU | 3.9 GiB | 15.0 GiB | Room to spare | |
| GeForce RTX 509032 GiB · GPU | 3.9 GiB | 31.0 GiB | Room to spare | |
| GeForce RTX 3060 12 GB12 GiB · GPU | 3.9 GiB | 11.0 GiB | Room to spare | |
| GeForce RTX 4060 Ti 16 GB16 GiB · GPU | 3.9 GiB | 15.0 GiB | Room to spare | |
| GeForce RTX 4070 Ti SUPER16 GiB · GPU | 3.9 GiB | 15.0 GiB | Room to spare | |
| GeForce RTX 309024 GiB · GPU | 3.9 GiB | 23.0 GiB | Room to spare | |
| GeForce RTX 3090 Ti24 GiB · GPU | 3.9 GiB | 23.0 GiB | Room to spare | |
| GeForce RTX 409024 GiB · GPU | 3.9 GiB | 23.0 GiB | Room to spare | |
| GeForce RTX 40608 GiB · GPU | 3.9 GiB | 7.0 GiB | Room to spare | |
| GeForce RTX 4060 Ti 8 GB8 GiB · GPU | 3.9 GiB | 7.0 GiB | Room to spare | |
| GeForce RTX 4070 GDDR6X12 GiB · GPU | 3.9 GiB | 11.0 GiB | Room to spare | |
| GeForce RTX 4070 SUPER12 GiB · GPU | 3.9 GiB | 11.0 GiB | Room to spare | |
| GeForce RTX 4070 Ti12 GiB · GPU | 3.9 GiB | 11.0 GiB | Room to spare | |
| GeForce RTX 408016 GiB · GPU | 3.9 GiB | 15.0 GiB | Room to spare | |
| GeForce RTX 4080 SUPER16 GiB · GPU | 3.9 GiB | 15.0 GiB | Room to spare | |
| GeForce RTX 5090 Laptop GPU 24 GB24 GiB · GPU | 3.9 GiB | 23.0 GiB | Room to spare | |
| GeForce RTX 5080 Laptop GPU 16 GB16 GiB · GPU | 3.9 GiB | 15.0 GiB | Room to spare | |
| GeForce RTX 5070 Ti Laptop GPU 12 GB12 GiB · GPU | 3.9 GiB | 11.0 GiB | Room to spare | |
| GeForce RTX 5070 Laptop GPU 12 GB12 GiB · GPU | 3.9 GiB | 11.0 GiB | Room to spare | |
| GeForce RTX 5070 Laptop GPU 8 GB8 GiB · GPU | 3.9 GiB | 7.0 GiB | Room to spare | |
| GeForce RTX 5060 Laptop GPU 8 GB8 GiB · GPU | 3.9 GiB | 7.0 GiB | Room to spare | |
| GeForce RTX 5050 Laptop GPU 8 GB8 GiB · GPU | 3.9 GiB | 7.0 GiB | Room to spare | |
| GeForce RTX 4090 Laptop GPU 16 GB16 GiB · GPU | 3.9 GiB | 15.0 GiB | Room to spare | |
| GeForce RTX 4080 Laptop GPU 12 GB12 GiB · GPU | 3.9 GiB | 11.0 GiB | Room to spare | |
| GeForce RTX 4070 Laptop GPU 8 GB8 GiB · GPU | 3.9 GiB | 7.0 GiB | Room to spare | |
| GeForce RTX 4060 Laptop GPU 8 GB8 GiB · GPU | 3.9 GiB | 7.0 GiB | Room to spare | |
| GeForce RTX 4050 Laptop GPU 6 GB6 GiB · GPU | 3.9 GiB | 5.0 GiB | Tight fit | |
| RTX PRO 2000 Blackwell16 GiB · GPU | 3.9 GiB | 15.0 GiB | Room to spare | |
| RTX PRO 4000 Blackwell SFF24 GiB · GPU | 3.9 GiB | 23.0 GiB | Room to spare | |
| RTX PRO 4000 Blackwell24 GiB · GPU | 3.9 GiB | 23.0 GiB | Room to spare | |
| RTX PRO 4500 Blackwell Workstation32 GiB · GPU | 3.9 GiB | 31.0 GiB | Room to spare | |
| RTX PRO 5000 Blackwell 48 GB48 GiB · GPU | 3.9 GiB | 47.0 GiB | Room to spare | |
| RTX PRO 5000 Blackwell 72 GB72 GiB · GPU | 3.9 GiB | 71.0 GiB | Room to spare | |
| RTX PRO 6000 Blackwell Workstation96 GiB · GPU | 3.9 GiB | 95.0 GiB | Room to spare | |
| RTX PRO 6000 Blackwell Max-Q96 GiB · GPU | 3.9 GiB | 95.0 GiB | Room to spare | |
| RTX A600048 GiB · GPU | 3.9 GiB | 47.0 GiB | Room to spare | |
| RTX 6000 Ada Generation48 GiB · GPU | 3.9 GiB | 47.0 GiB | Room to spare | |
| RTX 5000 Ada Generation32 GiB · GPU | 3.9 GiB | 31.0 GiB | Room to spare | |
| A100 80 GB PCIe80 GiB · GPU | 3.9 GiB | 78.0 GiB | Room to spare | |
| A100 80 GB SXM80 GiB · GPU | 3.9 GiB | 78.0 GiB | Room to spare | |
| H100 80 GB SXM80 GiB · GPU | 3.9 GiB | 78.0 GiB | Room to spare | |
| H100 NVL 94 GB PCIe94 GiB · GPU | 3.9 GiB | 92.0 GiB | Room to spare | |
| H200 141 GB SXM141 GiB · GPU | 3.9 GiB | 139.0 GiB | Room to spare | |
| H200 NVL 141 GB PCIe141 GiB · GPU | 3.9 GiB | 139.0 GiB | Room to spare | |
| L40S48 GiB · GPU | 3.9 GiB | 46.0 GiB | Room to spare | |
| B200 180 GB SXM180 GiB · GPU | 3.9 GiB | 178.0 GiB | Room to spare | |
| B300 288 GB SXM288 GiB · GPU | 3.9 GiB | 286.0 GiB | Room to spare | |
| A4048 GiB · GPU | 3.9 GiB | 46.0 GiB | Room to spare | |
| Radeon RX 680016 GiB · GPU | 3.9 GiB | 15.0 GiB | Room to spare | |
| Radeon RX 6800 XT16 GiB · GPU | 3.9 GiB | 15.0 GiB | Room to spare | |
| Radeon RX 6950 XT16 GiB · GPU | 3.9 GiB | 15.0 GiB | Room to spare | |
| Radeon RX 7600 XT16 GiB · GPU | 3.9 GiB | 15.0 GiB | Room to spare | |
| Radeon RX 7700 XT12 GiB · GPU | 3.9 GiB | 11.0 GiB | Room to spare | |
| Radeon RX 7800 XT16 GiB · GPU | 3.9 GiB | 15.0 GiB | Room to spare | |
| Radeon RX 7900 GRE16 GiB · GPU | 3.9 GiB | 15.0 GiB | Room to spare | |
| Radeon RX 7900 XT20 GiB · GPU | 3.9 GiB | 19.0 GiB | Room to spare | |
| Radeon RX 7900 XTX24 GiB · GPU | 3.9 GiB | 23.0 GiB | Room to spare | |
| Radeon RX 9060 XT 16 GB16 GiB · GPU | 3.9 GiB | 15.0 GiB | Room to spare | |
| Radeon RX 9070 GRE12 GiB · GPU | 3.9 GiB | 11.0 GiB | Room to spare | |
| Radeon RX 907016 GiB · GPU | 3.9 GiB | 15.0 GiB | Room to spare | |
| Radeon RX 9070 XT16 GiB · GPU | 3.9 GiB | 15.0 GiB | Room to spare | |
| Radeon AI PRO R970032 GiB · GPU | 3.9 GiB | 31.0 GiB | Room to spare | |
| Radeon PRO W7800 32 GB32 GiB · GPU | 3.9 GiB | 31.0 GiB | Room to spare | |
| Radeon PRO W7800 48 GB48 GiB · GPU | 3.9 GiB | 47.0 GiB | Room to spare | |
| Radeon PRO W790048 GiB · GPU | 3.9 GiB | 47.0 GiB | Room to spare | |
| Instinct MI210 PCIe64 GiB · GPU | 3.9 GiB | 62.0 GiB | Room to spare | |
| Instinct MI300X OAM192 GiB · GPU | 3.9 GiB | 190.0 GiB | Room to spare | |
| Instinct MI325X OAM256 GiB · GPU | 3.9 GiB | 254.0 GiB | Room to spare | |
| Instinct MI350X OAM288 GiB · GPU | 3.9 GiB | 286.0 GiB | Room to spare | |
| Instinct MI355X OAM288 GiB · GPU | 3.9 GiB | 286.0 GiB | Room to spare | |
| Instinct MI350P PCIe144 GiB · GPU | 3.9 GiB | 142.0 GiB | Room to spare | |
| Arc Pro B5016 GiB · GPU | 3.9 GiB | 15.0 GiB | Room to spare | |
| Arc Pro B6024 GiB · GPU | 3.9 GiB | 23.0 GiB | Room to spare | |
| Arc Pro B6532 GiB · GPU | 3.9 GiB | 31.0 GiB | Room to spare | |
| Arc Pro B7032 GiB · GPU | 3.9 GiB | 31.0 GiB | Room to spare | |
| Arc A770 16 GB16 GiB · GPU | 3.9 GiB | 15.0 GiB | Room to spare | |
| Arc B57010 GiB · GPU | 3.9 GiB | 9.0 GiB | Room to spare | |
| Arc B58012 GiB · GPU | 3.9 GiB | 11.0 GiB | Room to spare | |
| Mac mini M1 · 8 GB (8-core GPU)8 GiB · Unified memory | 3.9 GiB | 6.0 GiB | Room to spare | |
| Mac mini M1 · 16 GB (8-core GPU)16 GiB · Unified memory | 3.9 GiB | 12.0 GiB | Room to spare | |
| Mac mini M2 · 8 GB (10-core GPU)8 GiB · Unified memory | 3.9 GiB | 6.0 GiB | Room to spare | |
| Mac mini M2 · 16 GB (10-core GPU)16 GiB · Unified memory | 3.9 GiB | 12.0 GiB | Room to spare | |
| Mac mini M2 · 24 GB (10-core GPU)24 GiB · Unified memory | 3.9 GiB | 18.0 GiB | Room to spare | |
| Mac mini M2 Pro · 16 GB (16-core GPU)16 GiB · Unified memory | 3.9 GiB | 12.0 GiB | Room to spare | |
| Mac mini M2 Pro · 32 GB (16-core GPU)32 GiB · Unified memory | 3.9 GiB | 24.0 GiB | Room to spare | |
| Mac mini M2 Pro · 16 GB (19-core GPU)16 GiB · Unified memory | 3.9 GiB | 12.0 GiB | Room to spare | |
| Mac mini M2 Pro · 32 GB (19-core GPU)32 GiB · Unified memory | 3.9 GiB | 24.0 GiB | Room to spare | |
| MacBook Air M3 · 8 GB (8-core GPU)8 GiB · Unified memory | 3.9 GiB | 6.0 GiB | Room to spare | |
| MacBook Air M3 · 16 GB (8-core GPU)16 GiB · Unified memory | 3.9 GiB | 12.0 GiB | Room to spare | |
| MacBook Air M3 · 24 GB (8-core GPU)24 GiB · Unified memory | 3.9 GiB | 18.0 GiB | Room to spare | |
| MacBook Air M3 · 8 GB (10-core GPU)8 GiB · Unified memory | 3.9 GiB | 6.0 GiB | Room to spare | |
| MacBook Air M3 · 16 GB (10-core GPU)16 GiB · Unified memory | 3.9 GiB | 12.0 GiB | Room to spare | |
| MacBook Air M3 · 24 GB (10-core GPU)24 GiB · Unified memory | 3.9 GiB | 18.0 GiB | Room to spare | |
| MacBook Pro M1 Pro · 16 GB (16-core GPU)16 GiB · Unified memory | 3.9 GiB | 12.0 GiB | Room to spare | |
| MacBook Pro M1 Pro · 32 GB (16-core GPU)32 GiB · Unified memory | 3.9 GiB | 24.0 GiB | Room to spare | |
| Mac Studio M1 Max · 32 GB (24-core GPU)32 GiB · Unified memory | 3.9 GiB | 24.0 GiB | Room to spare | |
| Mac Studio M1 Max · 64 GB (24-core GPU)64 GiB · Unified memory | 3.9 GiB | 48.0 GiB | Room to spare | |
| Mac Studio M1 Max · 32 GB (32-core GPU)32 GiB · Unified memory | 3.9 GiB | 24.0 GiB | Room to spare | |
| Mac Studio M1 Max · 64 GB (32-core GPU)64 GiB · Unified memory | 3.9 GiB | 48.0 GiB | Room to spare | |
| Mac Studio M1 Ultra · 64 GB (48-core GPU)64 GiB · Unified memory | 3.9 GiB | 48.0 GiB | Room to spare | |
| Mac Studio M1 Ultra · 128 GB (48-core GPU)128 GiB · Unified memory | 3.9 GiB | 96.0 GiB | Room to spare | |
| Mac Studio M1 Ultra · 64 GB (64-core GPU)64 GiB · Unified memory | 3.9 GiB | 48.0 GiB | Room to spare | |
| Mac Studio M1 Ultra · 128 GB (64-core GPU)128 GiB · Unified memory | 3.9 GiB | 96.0 GiB | Room to spare | |
| Mac Studio M2 Max · 32 GB (30-core GPU)32 GiB · Unified memory | 3.9 GiB | 24.0 GiB | Room to spare | |
| Mac Studio M2 Max · 64 GB (30-core GPU)64 GiB · Unified memory | 3.9 GiB | 48.0 GiB | Room to spare | |
| Mac Studio M2 Max · 32 GB (38-core GPU)32 GiB · Unified memory | 3.9 GiB | 24.0 GiB | Room to spare | |
| Mac Studio M2 Max · 64 GB (38-core GPU)64 GiB · Unified memory | 3.9 GiB | 48.0 GiB | Room to spare | |
| Mac Studio M2 Max · 96 GB (38-core GPU)96 GiB · Unified memory | 3.9 GiB | 72.0 GiB | Room to spare | |
| Mac Studio M2 Ultra · 64 GB (60-core GPU)64 GiB · Unified memory | 3.9 GiB | 48.0 GiB | Room to spare | |
| Mac Studio M2 Ultra · 128 GB (60-core GPU)128 GiB · Unified memory | 3.9 GiB | 96.0 GiB | Room to spare | |
| Mac Studio M2 Ultra · 192 GB (60-core GPU)192 GiB · Unified memory | 3.9 GiB | 144.0 GiB | Room to spare | |
| Mac Studio M2 Ultra · 64 GB (76-core GPU)64 GiB · Unified memory | 3.9 GiB | 48.0 GiB | Room to spare | |
| Mac Studio M2 Ultra · 128 GB (76-core GPU)128 GiB · Unified memory | 3.9 GiB | 96.0 GiB | Room to spare | |
| Mac Studio M2 Ultra · 192 GB (76-core GPU)192 GiB · Unified memory | 3.9 GiB | 144.0 GiB | Room to spare | |
| MacBook Pro M3 Pro · 18 GB (18-core GPU)18 GiB · Unified memory | 3.9 GiB | 13.5 GiB | Room to spare | |
| MacBook Pro M3 Pro · 36 GB (18-core GPU)36 GiB · Unified memory | 3.9 GiB | 27.0 GiB | Room to spare | |
| MacBook Pro M3 Max · 36 GB (30-core GPU)36 GiB · Unified memory | 3.9 GiB | 27.0 GiB | Room to spare | |
| MacBook Pro M3 Max · 96 GB (30-core GPU)96 GiB · Unified memory | 3.9 GiB | 72.0 GiB | Room to spare | |
| MacBook Pro M3 Max · 48 GB (40-core GPU)48 GiB · Unified memory | 3.9 GiB | 36.0 GiB | Room to spare | |
| MacBook Pro M3 Max · 64 GB (40-core GPU)64 GiB · Unified memory | 3.9 GiB | 48.0 GiB | Room to spare | |
| MacBook Pro M3 Max · 128 GB (40-core GPU)128 GiB · Unified memory | 3.9 GiB | 96.0 GiB | Room to spare | |
| Mac mini M4 · 16 GB (10-core GPU)16 GiB · Unified memory | 3.9 GiB | 12.0 GiB | Room to spare | |
| Mac mini M4 · 24 GB (10-core GPU)24 GiB · Unified memory | 3.9 GiB | 18.0 GiB | Room to spare | |
| Mac mini M4 · 32 GB (10-core GPU)32 GiB · Unified memory | 3.9 GiB | 24.0 GiB | Room to spare | |
| Mac mini M4 Pro · 24 GB (16-core GPU)24 GiB · Unified memory | 3.9 GiB | 18.0 GiB | Room to spare | |
| Mac mini M4 Pro · 48 GB (16-core GPU)48 GiB · Unified memory | 3.9 GiB | 36.0 GiB | Room to spare | |
| Mac mini M4 Pro · 64 GB (16-core GPU)64 GiB · Unified memory | 3.9 GiB | 48.0 GiB | Room to spare | |
| Mac mini M4 Pro · 24 GB (20-core GPU)24 GiB · Unified memory | 3.9 GiB | 18.0 GiB | Room to spare | |
| Mac mini M4 Pro · 48 GB (20-core GPU)48 GiB · Unified memory | 3.9 GiB | 36.0 GiB | Room to spare | |
| Mac mini M4 Pro · 64 GB (20-core GPU)64 GiB · Unified memory | 3.9 GiB | 48.0 GiB | Room to spare | |
| Mac Studio M4 Max · 36 GB (32-core GPU)36 GiB · Unified memory | 3.9 GiB | 27.0 GiB | Room to spare | |
| Mac Studio M4 Max · 48 GB (40-core GPU)48 GiB · Unified memory | 3.9 GiB | 36.0 GiB | Room to spare | |
| Mac Studio M4 Max · 64 GB (40-core GPU)64 GiB · Unified memory | 3.9 GiB | 48.0 GiB | Room to spare | |
| Mac Studio M4 Max · 128 GB (40-core GPU)128 GiB · Unified memory | 3.9 GiB | 96.0 GiB | Room to spare | |
| Mac Studio M3 Ultra · 96 GB (60-core GPU)96 GiB · Unified memory | 3.9 GiB | 72.0 GiB | Room to spare | |
| Mac Studio M3 Ultra · 256 GB (60-core GPU)256 GiB · Unified memory | 3.9 GiB | 192.0 GiB | Room to spare | |
| Mac Studio M3 Ultra · 96 GB (80-core GPU)96 GiB · Unified memory | 3.9 GiB | 72.0 GiB | Room to spare | |
| Mac Studio M3 Ultra · 256 GB (80-core GPU)256 GiB · Unified memory | 3.9 GiB | 192.0 GiB | Room to spare | |
| Mac Studio M3 Ultra · 512 GB (80-core GPU)512 GiB · Unified memory | 3.9 GiB | 384.0 GiB | Room to spare | |
| MacBook Pro M5 · 16 GB (10-core GPU)16 GiB · Unified memory | 3.9 GiB | 12.0 GiB | Room to spare | |
| MacBook Pro M5 · 24 GB (10-core GPU)24 GiB · Unified memory | 3.9 GiB | 18.0 GiB | Room to spare | |
| MacBook Pro M5 · 32 GB (10-core GPU)32 GiB · Unified memory | 3.9 GiB | 24.0 GiB | Room to spare | |
| MacBook Pro M5 Pro · 24 GB (16-core GPU)24 GiB · Unified memory | 3.9 GiB | 18.0 GiB | Room to spare | |
| MacBook Pro M5 Pro · 48 GB (16-core GPU)48 GiB · Unified memory | 3.9 GiB | 36.0 GiB | Room to spare | |
| MacBook Pro M5 Pro · 24 GB (20-core GPU)24 GiB · Unified memory | 3.9 GiB | 18.0 GiB | Room to spare | |
| MacBook Pro M5 Pro · 48 GB (20-core GPU)48 GiB · Unified memory | 3.9 GiB | 36.0 GiB | Room to spare | |
| MacBook Pro M5 Pro · 64 GB (20-core GPU)64 GiB · Unified memory | 3.9 GiB | 48.0 GiB | Room to spare | |
| MacBook Pro M5 Max · 36 GB (32-core GPU)36 GiB · Unified memory | 3.9 GiB | 27.0 GiB | Room to spare | |
| MacBook Pro M5 Max · 48 GB (40-core GPU)48 GiB · Unified memory | 3.9 GiB | 36.0 GiB | Room to spare | |
| MacBook Pro M5 Max · 64 GB (40-core GPU)64 GiB · Unified memory | 3.9 GiB | 48.0 GiB | Room to spare | |
| MacBook Pro M5 Max · 128 GB (40-core GPU)128 GiB · Unified memory | 3.9 GiB | 96.0 GiB | Room to spare | |
| Ryzen AI Max+ 395 · 32 GB32 GiB · Unified memory | 3.9 GiB | 24.0 GiB | Room to spare | |
| Ryzen AI Max+ 395 · 64 GB64 GiB · Unified memory | 3.9 GiB | 48.0 GiB | Room to spare | |
| Ryzen AI Max+ 395 · 128 GB128 GiB · Unified memory | 3.9 GiB | 96.0 GiB | Room to spare | |
| Ryzen AI Max+ 392 · 32 GB32 GiB · Unified memory | 3.9 GiB | 24.0 GiB | Room to spare | |
| Ryzen AI Max+ 392 · 64 GB64 GiB · Unified memory | 3.9 GiB | 48.0 GiB | Room to spare | |
| Ryzen AI Max+ 392 · 128 GB128 GiB · Unified memory | 3.9 GiB | 96.0 GiB | Room to spare | |
| Ryzen AI Max+ 388 · 32 GB32 GiB · Unified memory | 3.9 GiB | 24.0 GiB | Room to spare | |
| Ryzen AI Max+ 388 · 64 GB64 GiB · Unified memory | 3.9 GiB | 48.0 GiB | Room to spare | |
| Ryzen AI Max+ 388 · 128 GB128 GiB · Unified memory | 3.9 GiB | 96.0 GiB | Room to spare | |
| NVIDIA DGX Spark · 128 GB128 GiB · Unified memory | 3.9 GiB | 96.0 GiB | Room to spare | |
| 2 × GeForce RTX 309048 GiB · Multi-GPU | 4.9 GiB | 46.0 GiB | Conditional · Room to spare | |
| 4 × GeForce RTX 309096 GiB · Multi-GPU | 6.9 GiB | 92.0 GiB | Conditional · Room to spare | |
| 2 × GeForce RTX 409048 GiB · Multi-GPU | 4.9 GiB | 46.0 GiB | Conditional · Room to spare | |
| 2 × GeForce RTX 509064 GiB · Multi-GPU | 4.9 GiB | 62.0 GiB | Conditional · Room to spare | |
| 4 × GeForce RTX 5090128 GiB · Multi-GPU | 6.9 GiB | 124.0 GiB | Conditional · Room to spare | |
| 2 × GeForce RTX 5060 Ti 16 GB32 GiB · Multi-GPU | 4.9 GiB | 30.0 GiB | Conditional · Room to spare | |
| 2 × RTX A600096 GiB · Multi-GPU | 4.9 GiB | 94.0 GiB | Conditional · Room to spare | |
| 2 × RTX PRO 5000 Blackwell 72 GB144 GiB · Multi-GPU | 4.9 GiB | 142.0 GiB | Conditional · Room to spare | |
| 2 × RTX PRO 6000 Blackwell Max-Q192 GiB · Multi-GPU | 4.9 GiB | 190.0 GiB | Conditional · Room to spare | |
| 4 × RTX PRO 6000 Blackwell Max-Q384 GiB · Multi-GPU | 6.9 GiB | 380.0 GiB | Conditional · Room to spare | |
| 2 × Radeon AI PRO R970064 GiB · Multi-GPU | 4.9 GiB | 62.0 GiB | Conditional · Room to spare | |
| 4 × Radeon AI PRO R9700128 GiB · Multi-GPU | 6.9 GiB | 124.0 GiB | Conditional · Room to spare | |
| 2 × Arc Pro B7064 GiB · Multi-GPU | 4.9 GiB | 62.0 GiB | Conditional · Room to spare | |
| 4 × Arc Pro B70128 GiB · Multi-GPU | 6.9 GiB | 124.0 GiB | Conditional · Room to spare | |
| 2 × Instinct MI210 PCIe128 GiB · Multi-GPU | 4.9 GiB | 124.0 GiB | Conditional · Room to spare | |
| 2 × H200 NVL 141 GB PCIe282 GiB · Multi-GPU | 4.9 GiB | 278.0 GiB | Conditional · Room to spare | |
| 4 × H200 NVL 141 GB PCIe564 GiB · Multi-GPU | 6.9 GiB | 556.0 GiB | Conditional · Room to spare | |
| GeForce RTX 3050 6 GB6 GiB · GPU | 3.9 GiB | 5.0 GiB | Tight fit | |
| GeForce RTX 3050 8 GB8 GiB · GPU | 3.9 GiB | 7.0 GiB | Room to spare | |
| GeForce RTX 3060 Ti 8 GB GDDR68 GiB · GPU | 3.9 GiB | 7.0 GiB | Room to spare | |
| GeForce RTX 30708 GiB · GPU | 3.9 GiB | 7.0 GiB | Room to spare | |
| GeForce RTX 3070 Ti8 GiB · GPU | 3.9 GiB | 7.0 GiB | Room to spare | |
| GeForce RTX 3080 10 GB10 GiB · GPU | 3.9 GiB | 9.0 GiB | Room to spare | |
| GeForce RTX 3080 12 GB12 GiB · GPU | 3.9 GiB | 11.0 GiB | Room to spare | |
| GeForce RTX 3080 Ti12 GiB · GPU | 3.9 GiB | 11.0 GiB | Room to spare | |
| Radeon RX 66008 GiB · GPU | 3.9 GiB | 7.0 GiB | Room to spare | |
| Radeon RX 6600 XT8 GiB · GPU | 3.9 GiB | 7.0 GiB | Room to spare | |
| Radeon RX 6650 XT8 GiB · GPU | 3.9 GiB | 7.0 GiB | Room to spare | |
| Radeon RX 6700 XT12 GiB · GPU | 3.9 GiB | 11.0 GiB | Room to spare | |
| Radeon RX 6750 XT12 GiB · GPU | 3.9 GiB | 11.0 GiB | Room to spare | |
| Arc A750 8 GB8 GiB · GPU | 3.9 GiB | 7.0 GiB | Room to spare | |
| Arc A770 8 GB8 GiB · GPU | 3.9 GiB | 7.0 GiB | Room to spare |
Architecture and calculation
Gated DeltaNet states use FP32; convolution states use the model dtype. Full-attention layers alone grow with text length. MTP speculation is excluded from cache.
- 8 full-attention layers × 4 KV heads × (256 key + 256 value dimensions).
24 recurrent layers maintain a fixed state per request, using FP32 state and 16-bit convolution buffers. They do not allocate a full attention cache for every token.
Precision is a storage estimate. Quantized files include scales and some higher-precision tensors; use the actual file size when known. A format also needs backend and hardware support. Full checkpoint count includes media/auxiliary weights when present; a text-only conversion can be smaller. Image/audio encoder working memory is not included. Full checkpoint count may include optional MTP or draft weights; cached speculative tokens are extra.
License: Apache 2.0. See the publisher’s terms.