Skip to content
Llama Builds

Alibaba Qwen · Released July 2025

Qwen3 235B-A22B (2507)

What it takes to run Qwen3 235B-A22B (2507) locally: memory by quantization, the GPUs that handle it, and how fast it runs.

Updated

Parameters
235B (22B active)
Architecture
Mixture of experts
Context
256K tokens
License
Apache 2.0 license
Memory at Q4_K_M
~149GB

What do you need to run Qwen3 235B-A22B (2507)?

Qwen3 235B-A22B (2507) needs about 149GB of VRAM at Q4_K_M with 8K context (142GB for the weights plus KV cache and overhead). Because it is a mixture-of-experts model with 22B active parameters, it generates text much faster than a dense model of the same size, and it tolerates partial CPU offload better.

See the full hardware guide for Qwen3 235B-A22B (2507)

What to buy

GPUs for Qwen3 235B-A22B (2507)

Unified memory

Ryzen AI Max+ PRO 495

192GB unified · ~$6,800

Quant
Q3_K_M
Speed (est.)
~11 tok/s
Runs well
Check price at Amazon

Multi-GPU

8× RTX 3090

192GB total · ~$8,400 used

Quant
Q4_K_M
Speed (est.)
~18 tok/s
Runs well
Check price at Amazon

Speeds are estimates at the quantization shown with 8K context. We may earn a commission when you buy through these links. Learn more.

Memory

Qwen3 235B-A22B (2507) VRAM requirements by quantization

Lower-bit quantizations shrink the model at some cost to quality. Q4_K_M is the usual sweet spot. Totals include the KV cache for the context length shown plus runtime overhead.

Memory needed for Qwen3 235B-A22B (2507) at each quantization level
QuantizationQualityWeightsTotal @ 8K ctxTotal @ 32K ctx
Q2_KRough88 GB93 GB98 GB
Q3_K_MUsable115 GB120 GB125 GB
Q4_K_MRecommendedDefault142 GB149 GB154 GB
Q5_K_MVery good167 GB175 GB179 GB
Q6_KNear-lossless194 GB202 GB207 GB
Q8_0Practically lossless250 GB259 GB264 GB
BF16Full precision470 GB486 GB491 GB

KV cache note: 2 x 94 layers x 4 KV heads x 128 dims x 2 bytes = 192,512 bytes (188 KiB) per token: about 1.6GB at 8K and 50GB at the full 262K context.

Performance

How fast Qwen3 235B-A22B (2507) runs

Single-user generation speed on popular GPUs that fit the model entirely in memory, at the best quantization each can hold.

  1. M5 Ultra 256GB

    Q4_K_M · 256GB

    ~27tok/s

  2. M5 Ultra 512GB

    Q8_0 · 512GB

    ~16tok/s

  3. M3 Ultra 512GB

    Q8_0 · 512GB

    ~11tok/s

  4. Ryzen AI Max+ PRO 495

    Q3_K_M · 192GB

    ~11tok/s

Estimated (method)

Overview

About Qwen3 235B-A22B (2507)

Qwen3 235B-A22B was Alibaba's open flagship for 2025, and its July 2025 (2507) Instruct and Thinking refreshes were among the best open models of that year. It has been superseded by the Qwen3.5 generation: Qwen3.5 122B-A10B matches or beats it in about half the memory, and Qwen3.5 397B-A17B replaced it at the top. At Q4_K_M it needs roughly 150GB with 8K context, which means a 192GB or larger Mac Studio, a multi-GPU workstation, or a server with lots of RAM for expert offload. With 22B active parameters it is noticeably slower per token than newer 10B to 17B active designs.

Strengths

  • +Broad knowledge and strong writing for an open model of its generation
  • +262K native context and separate Instruct and Thinking checkpoints
  • +Apache 2.0 license

Weaknesses

  • −Needs about 150GB at Q4_K_M, beyond any single consumer GPU
  • −Superseded by Qwen3.5 122B-A10B, which is about as capable in half the memory
  • −22B active parameters make it slower than newer sparse models on the same hardware
  • −Text only
ChatReasoningCodingAgents & toolsMultilingualWritingLong contextMathToggleable reasoning

Setup

How to run Qwen3 235B-A22B (2507)

Ollama
$ ollama run qwen3:235b
llama.cpp
$ llama-server -hf unsloth/Qwen3-235B-A22B-Instruct-2507-GGUF:Q4_K_M -c 8192 -ngl 99

Weights: Qwen/Qwen3-235B-A22B-Instruct-2507 · GGUF: unsloth/Qwen3-235B-A22B-Instruct-2507-GGUF

Alternatives

Models like Qwen3 235B-A22B (2507)

Alibaba Qwen

Qwen3.5 122B-A10B

122B (10B active) · 256K ctx

~77GBat Q4_K_M
ChatReasoningCoding

DeepSeek

DeepSeek V4 Flash

284B (13B active) · 1.0M ctx

~168GBat FP4
ReasoningCodingAgents & tools

Z.ai

GLM-5.3-Flash

320B (18B active) · 1.0M ctx

~201GBat Q4_K_M
CodingAgents & toolsReasoning

OpenAI

gpt-oss-120b

117B (5.1B active) · 128K ctx

~68GBat MXFP4
ReasoningCodingAgents & tools

Frequently asked questions

How much VRAM do you need to run Qwen3 235B-A22B (2507)?

About 149GB at Q4_K_M with 8K tokens of context: 142.5GB of weights, 1.6GB of KV cache, and 4.9GB of runtime overhead. Higher-precision quants and longer contexts need more; see the table above.

Is Qwen3 235B-A22B (2507) free for commercial use?

Qwen3 235B-A22B (2507) is released under the Apache 2.0 license. That is a permissive license, so commercial use is allowed.

How long a context does Qwen3 235B-A22B (2507) support?

Up to 262,144 tokens (256K). Long contexts need a lot of extra memory for the KV cache, so most local setups run 8K to 32K.

Sources: Qwen3 235B-A22B Instruct 2507 model card (Hugging Face), Qwen3 235B-A22B Thinking 2507 model card (Hugging Face). Capability score basis: Placed just below the gpt-oss-120b anchor from the Artificial Analysis Intelligence Index v3 standing of the 2507 Thinking variant (October 2025).