Qwen · 235.1B parameters · 22B active
Qwen3 235B-A22B: GPU memory requirements and which GPU to rent
Qwen3's flagship mixture-of-experts model: 235 billion parameters with 22 billion active per token, across 94 layers. Memory follows the total — about 235 GB at FP8, which one B300 or a multi-GPU node can hold.
- Weights at FP8235 GB470 GB at 16-bit · 129 GB at 4-bit
- FP8, 8k context, 1 sequence248 GBweights + KV cache + runtime
- Suggested here4× A100$2,596 a month · dedicated
Inference
How much GPU memory Qwen3 235B-A22B needs to serve
Total memory for one sequence at each context length, and for 16 concurrent sequences at 8k — weights, KV cache at 16-bit and runtime overhead included.
| Weights precision | Weights | 4k context | 32k context | 128k context | 8k × 16 sequences | Suggested here, 8k |
|---|---|---|---|---|---|---|
| 16-bit (BF16 / FP16) | 470 GB | 491 GB | 497 GB | 516 GB | 516 GB | 2× B300 · $7,580/mo |
| 8-bit (FP8 / INT8) | 235 GB | 247 GB | 252 GB | 271 GB | 271 GB | 4× A100 · $2,596/mo |
| 4-bit (GPTQ, AWQ, Q4) | 129 GB | 137 GB | 142 GB | 161 GB | 161 GB | 2× A100 · $1,298/mo |
GB are 109 bytes. Context lengths are shown for scale where the model supports them; an FP8 KV cache halves the cache part of each figure. As a mixture of experts, every expert stays resident, so memory follows the total parameter count, not the active one.
Hardware
GPU servers that run Qwen3 235B-A22B
For FP8 inference with an 8k context and one sequence: the configurations we would suggest, in order — one large card ahead of several small ones.
Fine-tuning Qwen3 235B-A22B
Memory for a 2k-token sequence and a micro-batch of 4, with gradient checkpointing. Full fine-tuning keeps gradients and Adam states for every parameter; LoRA and QLoRA train small adapters on a frozen base.
| Method | Memory | Suggested here |
|---|---|---|
| QLoRA, 4-bit base | 160 GB | 2× RTX PRO 6000 · $1,498/mo |
| LoRA, 16-bit base | 515 GB | 4× H200 · $6,760/mo |
| Full fine-tune, mixed precision | 3,795 GB | Multi-node |
Architecture at a glance
Grouped-query attention: every layer keeps keys and values for every token of the context. Figures from the model's published configuration; method on the calculator page.
Qwen3 235B-A22B hardware: questions answered
How much VRAM does Qwen3 235B-A22B need?
For the weights alone: about 470 GB at 16-bit, 235 GB at FP8 and 129 GB at 4-bit. Serving it at FP8 with an 8k-token context and one sequence takes about 248 GB in total once the KV cache and the runtime are counted; long contexts and many concurrent users add the KV cache on top — about 0.20 GB per 1,000 tokens at 16-bit.
Can Qwen3 235B-A22B run on a single GPU?
Yes, at FP8 with a moderate context: one NVIDIA B300 (288 GB HBM3e) holds it, for $3,790 a month here. Longer contexts or more concurrent users may need a larger card or a second GPU.
What is the cheapest GPU setup to run Qwen3 235B-A22B?
For FP8 inference with an 8k context and one sequence, the configuration we would suggest is 4× A100 · $2,596/mo; at 4-bit, 2× A100 · $1,298/mo. Suggestions favour one large card over several small ones: splitting a model across PCIe cards costs speed.
How much GPU memory do I need to fine-tune Qwen3 235B-A22B?
With a 2k-token sequence and a micro-batch of 4: about 160 GB for QLoRA on a 4-bit base, 515 GB for LoRA on a 16-bit base, and 3,795 GB for a full fine-tune with Adam in mixed precision. Gradient checkpointing is assumed; sharding the optimizer across GPUs spreads the full fine-tune over several cards.
Run Qwen3 235B-A22B on 4× A100 for $2,596 a month
Dedicated, single-tenant and live in under 10 minutes — with vLLM, the Hugging Face stack or Ollama preinstalled if you want. Paid in BTC, ETH or USDT, with no KYC.