Qwen3-235B-A22B VRAM requirements: every GGUF size and which GPUs fit

Qwen3-235B-A22B is the largest model in Qwen's original Qwen3 family, a mixture-of-experts (MoE) model released on Hugging Face on April 27, 2025 under the Apache 2.0 license. Hugging Face lists 235,093,634,560 parameters, and the model card says 22B are activated per token.^[1][5] Its config.json shows 94 layers, 128 experts with 8 picked per token, and grouped-query attention with 64 query heads and 4 KV heads of size 128.^[2] The model card gives a context of 32,768 tokens natively and 131,072 with YaRN rope scaling.^[5] Hugging Face counts about 351,183 downloads in the last month (checked November 2, 2026).^[1]

Because it is an MoE, only 22B parameters work on each token, so it generates much faster than a dense 235B model. But all of the weights still have to sit in memory (or be offloaded to system RAM).

Memory needed at each quantization

Below are the builds in Qwen's own GGUF repo (Q8_0, Q6_K, Q5_K_M, Q5_0, Q4_K_M) and Unsloth's GGUF repo (everything else, including BF16), using the exact file sizes on Hugging Face (1 GB = 1,000,000,000 bytes, multi-part files added together), largest first.^[3][4] Where both repos have the same quant, the sizes differ by about a kilobyte, so we list Qwen's. Builds whose names start with "UD-" are Unsloth's dynamic quants.

Which build should you pick? Q4_K_M is the usual balance of size and quality. With 256 GB or more, Q6_K or UD-Q6_K_XL. On a 128 GB Mac, Q3_K_M or UD-Q3_K_XL; treat 2-bit builds as a last resort.

Context length adds memory (the KV cache)

From config.json (94 layers, 4 KV heads, head size 128),^[2] an FP16 cache takes 2 × 94 × 4 × 128 × 2 bytes = 192,512 bytes per token, about 1.58 GB at 8K, 6.31 GB at 32K and 25.2 GB at 128K.

Contexts beyond 32,768 tokens need YaRN enabled; the model card shows the settings for llama.cpp, vLLM and SGLang and advises turning it on only when you need long inputs.^[5] Some engines can store the cache in 8-bit to roughly halve it.

Which GPU setups and Macs it fits on

Each cell compares the weights plus an 8K-token KV cache with the total memory of the setup.^{[6][7][8][9][10]}

Column headers show weights + an 8,192-token FP16 KV cache from config.json. Multi-GPU rows add up memory and assume the model is split evenly across the GPUs.

How to read it: "Fits" means weights + 8K cache use under 90% of memory. "Tight" means 90–100%: it may load, but the engine's own buffers or a longer context can push it over. "No" means it won't fit. The 90% line is our own rule of thumb, not a vendor figure. On a Mac, macOS also needs some of the unified memory, so treat those rows as optimistic.

Our recommendations

These follow from the file sizes and the 90% rule.^[3][4]

How it compares with other big Qwen models

4-bit sizes come from each model's Hugging Face GGUF repo, the same data as our comparison table.^[12]

Qwen3-235B-A22B is an April 2025 release; newer Qwen models on this site include Qwen3.5-122B-A10B (February 2026, about half the memory), Qwen3.8 Flash Next, Qwen3.8 35B-A3B and Qwen3.8 27B. Qwen also published July 2025 updates, Qwen3-235B-A22B-Instruct-2507 and -Thinking-2507. The Instruct-2507 config.json has the same 94 layers and 4 KV heads, so the KV cache numbers above apply to it, with a native context of 262,144 tokens.^[13] Its GGUF files are essentially the same size as the original's: in Unsloth's Instruct-2507 repo, Q4_K_M is 142.2 GB, UD-Q4_K_XL 134.3 GB, Q8_0 249.9 GB and BF16 470.3 GB (checked November 2, 2026).^[14] The difference is context: Instruct-2507 is non-thinking only and supports 262,144 tokens natively (extendable to about 1,010,000), so a full 262K-token FP16 KV cache adds about 50.5 GB, twice the 25.2 GB at 128K.^[13] Qwen's card suggests dropping to 32,768 tokens if you run out of memory.^[13] The original Qwen3-235B-A22B repo still has more downloads (351,183 in the last month vs 102,598 for Instruct-2507, checked November 2, 2026).^[1][13]

How to run it

With llama.cpp, LM Studio or Ollama, download a GGUF build from Qwen's or Unsloth's repo.^[3][4] The large builds are split into several files; llama.cpp loads the first part and finds the rest. If the model is bigger than your VRAM, llama.cpp can keep some experts in system RAM; see our MoE expert offload guide. With 22B active parameters, every token then waits on slower memory, so expect much lower speed. We have not measured its speed.

Cheapest GPUs to rent for Qwen3-235B-A22B

On-demand prices from Runpod's pricing page for setups that hold Q4_K_M with an 8K context (checked November 2, 2026; Community Cloud is cheaper but runs on shared hosts):^[11]

Prices checked on 2026-11-02. They change often, so confirm on the provider's page. Runpod lists per-GPU prices; the 2-GPU rows are that price times two. 160 GB holds Q4_K_M + 8K (143.7 GB) at 89.8%, right under our 90% line. Whether a multi-GPU pod is available at any moment varies. Rent on Runpod Referral link

See all GPU rental prices for more providers and multi-GPU setups, or try the VRAM calculator with your own quant and context.

Frequently asked questions

How much VRAM does Qwen3-235B-A22B need at 4-bit?

About 143.7 GB: the Q4_K_M weights (142.2 GB) plus an 8K-token FP16 KV cache (1.58 GB), before engine overhead.^[3][2] Unsloth's UD-Q4_K_XL is a little smaller at 134.1 GB.^[4]

Can I run Qwen3-235B-A22B on one GPU?

Not on a consumer GPU. The smallest build here, Q2_K, is 85.7 GB, more than any 24 GB or 32 GB card.^[4] A single 141 GB H200 holds Q3_K_M; otherwise you need two or more 80–96 GB GPUs, a Mac with 128 GB or more, or expert offloading to system RAM.

Will it run on a 128 GB Mac?

At 3 bits, yes by our rule: Q3_K_M (112.4 GB) with an 8K context uses 89% of 128 GB, and UD-Q3_K_XL (103.7 GB) leaves more room. IQ4_XS (125.5 GB) is tight and Q4_K_M does not fit.^[4][10] macOS keeps part of unified memory for itself by default, so the 3-bit builds may need the GPU memory limit raised; see our Mac guide.

How much does a 128K context add?

At FP16 the KV cache for 131,072 tokens is about 25.2 GB.^[2] That length needs YaRN enabled.^[5]

Is it open source?

The weights are released under the Apache 2.0 license, which allows commercial use.^[1]

How we calculated

• Weights are the exact GGUF file sizes on Hugging Face, checked November 2, 2026, with multi-part files added up.^[3][4]
• KV cache is calculated from config.json. We assume an FP16 cache.^[2]
• Device memory is from NVIDIA and Apple spec pages. Multi-GPU rows add up memory and assume an even split.
• Rental prices are the listed on-demand Runpod per-GPU rates from our GPU price data, checked November 2, 2026, times the GPU count.^[11]
• Not included: engine working buffers, the OS and display. These vary by engine and settings, so leave headroom.

Sources

All sources checked on 2026-11-02.

1. Hugging Face, Qwen/Qwen3-235B-A22B (parameter count, license, repo date, 30-day downloads; checked November 2, 2026). https://huggingface.co/Qwen/Qwen3-235B-A22B
2. Hugging Face, Qwen/Qwen3-235B-A22B config.json (94 layers, 4 KV heads, head size 128, 128 experts, 8 per token). https://huggingface.co/Qwen/Qwen3-235B-A22B/blob/main/config.json
3. Hugging Face, Qwen/Qwen3-235B-A22B-GGUF file list (Qwen's official GGUF sizes; checked November 2, 2026). https://huggingface.co/Qwen/Qwen3-235B-A22B-GGUF/tree/main
4. Hugging Face, unsloth/Qwen3-235B-A22B-GGUF file list (checked November 2, 2026). https://huggingface.co/unsloth/Qwen3-235B-A22B-GGUF/tree/main
5. Qwen3-235B-A22B model card (235B total / 22B activated, context length, YaRN settings). https://huggingface.co/Qwen/Qwen3-235B-A22B/blob/main/README.md
6. NVIDIA, A100 (80 GB). https://www.nvidia.com/en-us/data-center/a100/
7. NVIDIA, H100 (80 GB). https://www.nvidia.com/en-us/data-center/h100/
8. NVIDIA, H200 (141 GB). https://www.nvidia.com/en-us/data-center/h200/
9. NVIDIA, RTX PRO 6000 Blackwell (96 GB). https://www.nvidia.com/en-us/products/workstations/professional-desktop-gpus/rtx-pro-6000/
10. Apple, Mac Studio tech specs (128 GB, 256 GB and 512 GB unified memory options). https://www.apple.com/mac-studio/specs/
11. Runpod, GPU pricing (per-GPU prices; checked November 2, 2026). https://www.runpod.io/pricing
12. VRAMNeeded comparison table (sources listed per model). https://vramneeded.com/compare/
13. Hugging Face, Qwen/Qwen3-235B-A22B-Instruct-2507 config.json and model card (262,144-token native context, non-thinking only, OOM advice, 30-day downloads). https://huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507
14. Hugging Face, unsloth/Qwen3-235B-A22B-Instruct-2507-GGUF file list (checked November 2, 2026). https://huggingface.co/unsloth/Qwen3-235B-A22B-Instruct-2507-GGUF/tree/main