Free tools

Qwen 3.5 235B GPU requirements

What it takes to run Qwen 3.5 235B (MoE) locally: memory by quantization, the smallest GPU that fits, and the managed alternative.

FP16521 GB needed at 8K contextMulti-GPU
8-bit260 GB needed at 8K contextMac M3 Ultra
4-bit143 GB needed at 8K contextB200
144 GBmemory needed · Qwen 3.5 235B at 4-bit, 16K context

Smallest single device that fits: B200 (192 GB)

RTX 4090 (24 GB)7x needed
RTX 5090 (32 GB)5x needed
L40S (48 GB)4x needed
A100 (80 GB)2x needed
H100 (80 GB)2x needed
H200 (141 GB)2x needed
B200 (192 GB)Fits
Mac M4 Max (128 GB unified)2x needed
Mac M3 Ultra (512 GB unified)Fits

Or skip the hardware: run Qwen 3.5 235B on Allocate, token-metered.

How it works

1
Check the table
Qwen 3.5 235B at 8K context across FP16, 8-bit, and 4-bit, with the smallest single device that fits each.
2
Tune the calculator
Longer context grows the KV cache. The slider shows exactly how much memory that adds.
3
Decide how to run it
MoE weights must fit in memory in full, so serving this locally means a multi-GPU fleet, not one card.

Common questions

How much VRAM does Qwen 3.5 235B need?

At 8K context: roughly 521 GB at FP16, 260 GB at 8-bit, and 143 GB at 4-bit, including KV cache and runtime overhead. Longer context adds memory linearly.

Can a single GPU run Qwen 3.5 235B?

At 4-bit, yes: a B200 (192 GB) or larger handles it at 8K context. At FP16 you need multiple devices.

Why does Qwen 3.5 235B need so much memory as a MoE model?

Only some experts activate per token, which sets speed, but all 235 billion parameters must sit in memory. MoE trades memory for throughput.

What license is Qwen 3.5 235B under?

Apache 2.0. Permissive: run it, fine-tune it, and own the result.

Is there an alternative to buying GPUs for Qwen 3.5 235B?

Yes: managed inference. Allocate serves Qwen 3.5 235B token-metered inside your own boundary, and if you fine-tune it on your data, the weights belong to you. Idle time costs nothing.

More free tools

Allocate serves Qwen 3.5 235B token-metered inside your own boundary. No hardware to buy, and if you fine-tune it on your data, the weights are yours.