SAVRN Model Hub · Legend_ocr_qwen3-highavg
Legend_ocr_qwen3-highavg GPU Requirements
Legend_ocr_qwen3-highavg needs about 21 GB of GPU memory at 16-bit, 10.5 GB at 8-bit and 5.3 GB at 4-bit: its 8.8B parameters plus a 20% working margin. The cheapest setup at 16-bit is 1x MI300X from $1.85 an hour, about $1,350 a month around the clock. Every one of the 9 accelerators the SAVRN Index prices holds it on one card at 16-bit.
Every Accelerator, Every Precision
How many cards of each accelerator the SAVRN Index prices it takes to hold Legend_ocr_qwen3-highavg, and what that many cards cost an hour at the lowest listed on-demand price. Memory needed: 21 GB at 16-bit, 10.5 GB at 8-bit, 5.3 GB at 4-bit.
| Accelerator | Memory per card | Lowest price per card | 16-bit | 8-bit | 4-bit |
|---|---|---|---|---|---|
| H100 Voltage Park |
80 GB | $1.99 | 1 card $1.99/hr |
1 card $1.99/hr |
1 card $1.99/hr |
| H200 GMI Cloud |
141 GB | $2.60 | 1 card $2.60/hr |
1 card $2.60/hr |
1 card $2.60/hr |
| B200 Vultr |
180 GB | $3.50 | 1 card $3.50/hr |
1 card $3.50/hr |
1 card $3.50/hr |
| GB200 NVL72 GMI Cloud |
186 GB | $8.00 | 1 card $8.00/hr |
1 card $8.00/hr |
1 card $8.00/hr |
| MI300X Vultr |
192 GB | $1.85 | 1 card $1.85/hr |
1 card $1.85/hr |
1 card $1.85/hr |
| MI325X Vultr |
256 GB | $2.00 | 1 card $2.00/hr |
1 card $2.00/hr |
1 card $2.00/hr |
| B300 Massed Compute |
268 GB | $6.60 | 1 card $6.60/hr |
1 card $6.60/hr |
1 card $6.60/hr |
| GB300 NVL72 Verda |
279 GB | $9.06 | 1 card $9.06/hr |
1 card $9.06/hr |
1 card $9.06/hr |
| MI355X Vultr |
288 GB | $2.59 | 1 card $2.59/hr |
1 card $2.59/hr |
1 card $2.59/hr |
Running It Around the Clock
| Precision | Cheapest setup | Per hour | Per month (730 hours) |
|---|---|---|---|
| 16-bit | 1x MI300X (Vultr) | $1.85 | $1,350 |
| 8-bit | 1x MI300X (Vultr) | $1.85 | $1,350 |
| 4-bit | 1x MI300X (Vultr) | $1.85 | $1,350 |
One copy of the model on rented cards, busy or idle. Serving more users at once takes more copies or more memory for their contexts.
Memory at Longer Context
Every token in a sequence keeps a key and a value in every layer. From Legend_ocr_qwen3-highavg's published configuration, that cache adds this much at 16-bit for one sequence:
| Context | Key/value cache | Total with weights | Cheapest setup |
|---|---|---|---|
| 4,096 tokens | 0.6 GB | 21.6 GB | 1x MI300X $1.85/hr |
| 32,768 tokens | 4.8 GB | 25.8 GB | 1x MI300X $1.85/hr |
| 131,072 tokens | 19.3 GB | 40.3 GB | 1x MI300X $1.85/hr |
| 262,144 tokens (full) | 38.7 GB | 59.7 GB | 1x MI300X $1.85/hr |
An estimate from layers, key/value heads and head size, assuming full attention in every layer. Runtimes that quantize or page the cache use less.
Questions
How much VRAM does Legend_ocr_qwen3-highavg need?
About 21 GB at 16-bit; about 10.5 GB at 8-bit; about 5.3 GB at 4-bit: the weights plus 20% for the runtime and a short context. A long context needs more.
What is the cheapest GPU setup to run Legend_ocr_qwen3-highavg?
At 16-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand price the SAVRN Index lists.
Can Legend_ocr_qwen3-highavg run on a single H100?
Yes at 16-bit, 8-bit, 4-bit: an H100 has 80 GB and Legend_ocr_qwen3-highavg needs 21 GB at 16-bit.
How much memory does Legend_ocr_qwen3-highavg need at its full context length?
About 59.7 GB at 16-bit for one 262,144-token sequence: 21 GB for the weights and margin plus 38.7 GB of key/value cache, estimated from its published configuration.
Memory is the weights at that precision plus 20% for the runtime and a short context. Prices are the lowest on-demand hourly rates in the SAVRN Index, read Sep 24, 2026. Setups beyond eight cards, one server, are not listed.