SAVRN Model Hub · Qwen3-235B-A22B-Thinking-2507
Qwen3-235B-A22B-Thinking-2507 GPU Requirements
Qwen3-235B-A22B-Thinking-2507 needs about 564 GB of GPU memory at 16-bit, 282 GB at 8-bit and 141 GB at 4-bit: its 235.1B parameters plus a 20% working margin. The cheapest setup at 16-bit is 2x MI355X from $5.18 an hour, about $3,781 a month around the clock. None 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 Qwen3-235B-A22B-Thinking-2507, and what that many cards cost an hour at the lowest listed on-demand price. Memory needed: 564 GB at 16-bit, 282 GB at 8-bit, 141 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 | 8 cards $15.92/hr |
4 cards $7.96/hr |
2 cards $3.98/hr |
| H200 GMI Cloud |
141 GB | $2.60 | 5 cards $13.00/hr |
3 cards $7.80/hr |
2 cards $5.20/hr |
| B200 Vultr |
180 GB | $3.50 | 4 cards $14.00/hr |
2 cards $7.00/hr |
1 card $3.50/hr |
| GB200 NVL72 GMI Cloud |
186 GB | $8.00 | 4 cards $32.00/hr |
2 cards $16.00/hr |
1 card $8.00/hr |
| MI300X Vultr |
192 GB | $1.85 | 3 cards $5.55/hr |
2 cards $3.70/hr |
1 card $1.85/hr |
| MI325X Vultr |
256 GB | $2.00 | 3 cards $6.00/hr |
2 cards $4.00/hr |
1 card $2.00/hr |
| B300 Massed Compute |
268 GB | $6.60 | 3 cards $19.80/hr |
2 cards $13.20/hr |
1 card $6.60/hr |
| GB300 NVL72 Verda |
279 GB | $10.32 | 3 cards $30.96/hr |
2 cards $20.64/hr |
1 card $10.32/hr |
| MI355X Vultr |
288 GB | $2.59 | 2 cards $5.18/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 | 2x MI355X (Vultr) | $5.18 | $3,781 |
| 8-bit | 1x MI355X (Vultr) | $2.59 | $1,891 |
| 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 Qwen3-235B-A22B-Thinking-2507'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 | 0.8 GB | 565 GB | 2x MI355X $5.18/hr |
| 32,768 tokens | 6.3 GB | 571 GB | 2x MI355X $5.18/hr |
| 131,072 tokens | 25.2 GB | 589 GB | 3x MI325X $6.00/hr |
| 262,144 tokens (full) | 50.5 GB | 615 GB | 3x MI325X $6.00/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 Qwen3-235B-A22B-Thinking-2507 need?
About 564 GB at 16-bit; about 282 GB at 8-bit; about 141 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 Qwen3-235B-A22B-Thinking-2507?
At 16-bit, 2x MI355X from $5.18 an hour, at the lowest on-demand price the SAVRN Index lists.
Can Qwen3-235B-A22B-Thinking-2507 run on a single H100?
No. An H100 has 80 GB, and Qwen3-235B-A22B-Thinking-2507 needs 141 GB even at 4-bit, so it takes more than one card.
How much memory does Qwen3-235B-A22B-Thinking-2507 need at its full context length?
About 615 GB at 16-bit for one 262,144-token sequence: 564 GB for the weights and margin plus 50.5 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 Oct 7, 2026. Setups beyond eight cards, one server, are not listed.