Qwen 3 235B-A22B
By Alibaba · China
Updated 2026-07-13
Overview
Alibaba's flagship MoE — 235B total, 22B active per token across 128 experts. Hits 85.7 on AIME 2024 and 70.7 on LiveCodeBench, putting it in frontier-open territory.
When to pick this model
- You're running multi-GPU or a high-memory Apple Silicon machine and want frontier-open performance
- You need top-tier math and code reasoning under an Apache license
- You want MoE-class throughput (22B active) rather than dense 200B+ latency
- You're evaluating against closed frontier models and need a serious local baseline
VRAM requirements by quantization
| Quantization | VRAM required |
|---|---|
| Q4_K_M (recommended) | 142 GB |
| Q5_K_M | 170 GB |
| Q8_0 | 250 GB |
| FP16 (no quantization) | 470 GB |
VRAM figures include model weights plus a typical 8k KV cache and ~600 MB runtime overhead (Ollama / llama.cpp baseline). Add headroom for higher context lengths.
In practice, Qwen 3 235B-A22B is server-class even at Q4_K_M (142 GB). Stepping up to Q8_0 nearly doubles the footprint to 250 GB, and unquantized FP16 weights take 470 GB — plan your GPU around the Q4 or Q5 figure unless you specifically need the higher fidelity.
Without a GPU, Qwen 3 235B-A22B needs roughly 160 GB of system RAM to run on CPU via llama.cpp or Ollama — workable for background jobs, but far slower than GPU inference. Throughput estimates from our compatibility engine: around 3 tokens/sec on entry-level GPUs, on the order of 12 tokens/sec on a mid-range card, and up to 28 tokens/sec on high-end hardware — assuming the chosen quantization fully fits in VRAM.
What hardware do you need
The table below matches Qwen 3 235B-A22B to common GPU memory tiers, using the highest-fidelity quantization that fully fits each card class. Spilling layers to system RAM works but costs most of the speed, so size your card to the quantization you actually want to run.
| GPU memory | Example cards | Best fit for Qwen 3 235B-A22B |
|---|---|---|
| 8 GB | RTX 5070 Laptop, RTX 5060, RTX 5060 Ti 8GB | Does not fit — needs 142 GB at Q4_K_M |
| 12 GB | RTX 5070, RTX 5070 Ti Laptop, RTX 4080 Laptop | Does not fit — needs 142 GB at Q4_K_M |
| 16 GB | RTX 5080, RTX 4080 Super, Radeon RX 9070 XT | Does not fit — needs 142 GB at Q4_K_M |
| 24 GB | RTX 4090, Radeon RX 7900 XTX, RTX 5090 Laptop | Does not fit — needs 142 GB at Q4_K_M |
| 32 GB | RTX 5090 | Does not fit — needs 142 GB at Q4_K_M |
Which GPU should you buy to run Qwen 3 235B-A22B?
To run Qwen 3 235B-A22B locally at Q4, you need ~142 GB of VRAM. The best value for this is a Apple Mac Studio (64+ GB unified memory).
As an Amazon Associate, BestLLMfor earns from qualifying purchases, at no extra cost to you. It does not influence our independent rankings.
Published benchmark scores
| Benchmark | Score |
|---|---|
| AIME 2024 | 85.7 |
| AIME 2025 | 81.5 |
| LiveCodeBench v5 | 70.7 |
Scores published by the model author or aggregated from public leaderboards. Re-measured monthly by our editorial team.
To put Qwen 3 235B-A22B in context: its AIME 2024 score of 85.7 ranks #1 of the 8 catalog models with a published AIME 2024 result (catalog median 71.7). Scores are author-published and measured under different harnesses, so treat gaps of a few points as noise rather than a verdict.
Strengths
- Frontier-open scores on AIME 2024 (85.7) and LiveCodeBench (70.7)
- Only 22B active parameters — fast for its total size
- Instruct-2507 and Thinking-2507 variants available
- Apache 2.0
Limitations
- ~142GB at Q4 — needs multi-GPU or a 192GB+ Apple Silicon host
- Not realistic for laptop or single-GPU deployment
Typical workloads
In our catalog grid, Qwen 3 235B-A22B is filed under Frontier Reasoning, Advanced Code, Agents — the use cases where its size/quality trade-off makes the most sense. Its tags translate to concrete workloads: multi-step reasoning and math-flavoured tasks; multilingual workloads.
The 128k-token context window is large enough to hold entire codebases' worth of files or long reports in a single prompt, which is what makes local RAG and document analysis practical. The Apache 2.0 license is permissive, so shipping it inside a commercial product raises no special legal questions.
Architecture & training
Architecture: MoE · 128 experts, 8 active · 94 layers · GQA 64Q/4KV
Training: 36T tokens. Instruct-2507 and Thinking-2507 variants (July 2025).
Pick this when you have the hardware for frontier-open performance under an Apache license.
Quick start
ollama run qwen3:235bOr use the open-source MCP server to query this model from Claude Desktop, Cursor, or any MCP-compatible client.
Similar models worth comparing
Frequently asked questions
How much VRAM does Qwen 3 235B-A22B need?
At the recommended Q4_K_M quantization, Qwen 3 235B-A22B needs about 142 GB of VRAM. Q8_0 takes 250 GB, and unquantized FP16 weights take 470 GB.
Can Qwen 3 235B-A22B run without a GPU?
Yes — with roughly 160 GB of system RAM it runs CPU-only through llama.cpp or Ollama. Expect a fraction of GPU speed, which is fine for background or batch jobs but slow for interactive chat.
What context window does Qwen 3 235B-A22B support?
Qwen 3 235B-A22B supports a 128k-token context window (131,072 tokens).
Can I use Qwen 3 235B-A22B commercially?
Yes. Qwen 3 235B-A22B is released under Apache 2.0, a permissive open-source license that allows commercial use, modification and redistribution.
How fast is Qwen 3 235B-A22B on consumer hardware?
Our compatibility engine estimates on the order of 12 tokens/sec on a mid-range GPU and up to 28 tokens/sec on high-end cards, assuming the quantization fully fits in VRAM.
Which quantization of Qwen 3 235B-A22B should I download first?
Start with Q4_K_M (142 GB) — the standard size/quality sweet spot. Step up to Q5_K_M or Q8_0 only if you have VRAM headroom. It does not fit a single 24 GB consumer card — plan for multi-GPU or server hardware.