ai.canshu.com
中文
VRAM Calculator Advisor & TCO API Pricing Compare Models Leaderboard GPUs

Search models, GPUs, and tools

↑↓ Navigate · Enter Open · ★ Watchlist prioritized ai.canshu.com
Alibaba Qwen
Open-Weight Alibaba Qwen · MoE 235.1B / A22.2B · Text · BF16

Qwen3-235B-A22B (MoE) Architecture Parameters, Benchmarks & VRAM

235B total / 22B active MoE flagship suited for high-memory Mac Studios and multi-GPU servers.

🧬 Qwen3-235B-A22B (MoE) Architecture & Tensor Specifications

Verified directly against HuggingFace (Qwen/Qwen3-235B-A22B) config.json & Model Card

HF Class: Qwen3MoeForCausalLM

1. Backbone & Dimensions

qwen3_moe
Total / Active Params
235.1B / 22.2B
Non-Embedding Params
234B
Hidden Size (d_model)
4,096
FFN Intermediate Size
12,288
MoE Expert Dim
1,536
MoE Experts
128 Routed · Top-8
Activation / Norm
SiLU (SwiGLU) · ε=1e-6
Native Tensor Dtype
BF16

2. Attention & Layer Layout

GQA
Num Hidden Layers
94 Layers
Hidden Layer Layout
94-Layer Flagship MoE (128 Experts, 8 Activated)
Attention Heads (Q / KV)
64 Q / 4 KV (16:1 GQA)
Head Dimension
128

3. Context, RoPE & Tokenizer

Tokenizer
Native Context Length
32,768 tokens
Max Extended Context
131,072 tokens
RoPE Configuration
RoPE (θ=1,000,000 · 原生 32K · YaRN 4x→128K)
Vocab Size
151,936 (Untied)
Special Token IDs
BOS: 151643 · EOS: [151645,151643] · PAD: 151643
Language Support
119 languages & dialects

📋 Qwen3-235B-A22B (MoE) Official API & System Card Specifications

Verified against official Alibaba Qwen developer documentation & pricing

API Model ID: qwen3-235b-a22b

1. Identity & Capabilities

API Spec
API Model ID
qwen3-235b-a22b
Release Date
2025-04-29
Weight License
Open-Weight (API Spec)
Native Reasoning
✅ Supported
Function Calling & JSON Schema
✅ 原生支持
Batch API 异步批处理
✅ Supported (50% off)

2. Context, Output & Modalities

Limits
Max Context Window
128,000 tokens (128K)
Max Output Tokens
16,000 tokens
Input Modalities
Text
Output Modalities
Text
Direct China Access
✅ Direct CN + RMB
Deployment Mode
Open-Weight + Cloud API

3. Official Pricing & Cache

CNY / 1M
Input / 1M Tokens
¥ 2
Cached Input / 1M
¥ 0.4 -80%
Output / 1M Tokens
¥ 8
Tiered Pricing Notes
Alibaba Bailian: Hosted on Alibaba Cloud Bailian (CN Direct)

🏆 Official Benchmark Results

Sourced from official System Card / Model Card & technical reports

Arena Elo ~1472 AA Index 47
📚 Knowledge Thinking
MMLU-Pro 82.8 %

A more robust and challenging multi-task language understanding benchmark that extends MMLU by expanding multiple-choice options from 4 to 10, eliminating trivial questions, and focusing on reasoning-intensive tasks.

📚 Knowledge Thinking
MMLU-Redux 92.7 %

An improved version of the MMLU benchmark featuring manually re-annotated questions to identify and correct errors in the original dataset. Provides more reliable evaluation metrics for language models by addressing dataset quality issues found in the original MMLU.

🧠 Reasoning Thinking
GPQA Diamond 71.1 %

PhD-level Google-proof science reasoning questions

🧠 Reasoning Thinking
MATH-500 98 %

MATH-500 is a subset of the MATH dataset containing 500 challenging competition mathematics problems from AMC 10, AMC 12, AIME, and other mathematics competitions.

🧠 Reasoning Thinking
AIME 2024 85.7 %

American Invitational Mathematics Examination 2024, consisting of 30 challenging mathematical reasoning problems from AIME I and AIME II competitions.

🧠 Reasoning Thinking
AIME 2025 81.5 %

All 30 problems from the 2025 American Invitational Mathematics Examination (AIME I and AIME II), testing olympiad-level mathematical reasoning with integer answers from 000-999.

💻 Coding Thinking
LiveCodeBench v5 70.7 %

LiveCodeBench is a holistic and contamination-free evaluation benchmark for large language models for code.

🤖 Agentic Thinking
BFCL v3 70.8 %

Berkeley Function Calling Leaderboard v3 (BFCL-v3) is an advanced benchmark that evaluates large language models' function calling capabilities through multi-turn and multi-step interactions.

🎛️ Official Sampling & Reasoning Parameters

Recommended generation settings & API parameter constraints

Thinking Mode (Reasoning & Math/Code)
temperature=0.6 top_p=0.95 top_k=20 min_p=0 max_tokens=32,768

💡 Do NOT use greedy decoding (temp=0); use 38,912 max tokens for competition math/code.

Non-Thinking Mode (General Chat)
temperature=0.7 top_p=0.8 top_k=20 min_p=0 presence_penalty=1.5 max_tokens=32,768

💡 Recommended for fast general chat; presence_penalty 0~2 reduces repetition.

📊 Qwen3-235B-A22B (MoE) Full Benchmark Scorecard

21 public benchmarks · ranks among 154 models · synced 2026-11-20

⚔️ Compare head-to-head

📐 Math 12 items

GSM8k #2/4 94.4%

Grade School Math 8K, a dataset of 8.5K high-quality linguistically diverse grade school math word problems requiring multi-step reasoning and elementary arithmetic operations.

BBH #1/2 88.9%

Big-Bench Hard (BBH) is a suite of 23 challenging tasks selected from BIG-Bench for which prior language model evaluations did not outperform the average human-rater.

MMLU #3/11 87.8%

Massive Multitask Language Understanding benchmark testing knowledge across 57 diverse subjects including STEM, humanities, social sciences, and professional domains.

MMLU-Redux #14/15 87.4%

An improved version of the MMLU benchmark featuring manually re-annotated questions to identify and correct errors in the original dataset. Provides more reliable evaluation metrics for language models by addressing dataset quality issues found in the original MMLU.

MMMLU #12/20 86.7%

Multilingual Massive Multitask Language Understanding dataset released by OpenAI, featuring professionally translated MMLU test questions across 14 languages including Arabic, Bengali, German, Spanish, French, Hindi, Indonesian, Italian, Japanese, Korean, Portuguese, Swahili…

AIME 2024 #8/14 85.7%

American Invitational Mathematics Examination 2024, consisting of 30 challenging mathematical reasoning problems from AIME I and AIME II competitions.

MGSM #5/7 83.5%

MGSM (Multilingual Grade School Math) is a benchmark of grade-school math problems. Contains 250 grade-school math problems manually translated from the GSM8K dataset into ten typologically diverse languages: Spanish, French, German, Russian, Chinese, Japanese, Thai, Swahili…

AIME 2025 #21/29 81.5%

All 30 problems from the 2025 American Invitational Mathematics Examination (AIME I and AIME II), testing olympiad-level mathematical reasoning with integer answers from 000-999.

LiveBench #7/27 77.1%

LiveBench is a challenging, contamination-limited LLM benchmark that addresses test set contamination by releasing new questions monthly based on recently-released datasets, arXiv papers, news articles, and IMDb movie synopses.

MATH #5/9 71.8%

MATH dataset contains 12,500 challenging competition mathematics problems from AMC 10, AMC 12, AIME, and other mathematics competitions.

MMLU-Pro #36/39 68.2%

A more robust and challenging multi-task language understanding benchmark that extends MMLU by expanding multiple-choice options from 4 to 10, eliminating trivial questions, and focusing on reasoning-intensive tasks.

SuperGPQA #16/16 44.1%

SuperGPQA is a comprehensive benchmark that evaluates large language models across 285 graduate-level academic disciplines.

Source: ZeroEval / llm-stats.com (✓ = third-party verified, others vendor-reported).

⚡ Qwen3-235B-A22B (MoE) Interactive VRAM & Token Speed Calculator

Concurrency (Batch):
MoE 235.1B / A22.2B·94L (d=4096)·64Q/4KV·Ctx 128K
Runnable Models →
512 GB Unified·1200 GB/s·Safe Usable 460.8 GB
Smooth Fit148.98 / 460.8 GB~52.7 tok/s
3. Weight Quantization
4. Per-Request Context16K (16,384 tok)
Verified: 2026-11-20·Roofline bandwidth/compute estimate (Batch=1)
Smooth Fit
148.98/ 460.8 GB usable

Physical 512 GB · Load 32.3%

Theoretical Decode Speed
~52.7tokens/s

MoE: reads only 13.46 GB active weights/step

Max Context at Batch=1
128K/ 128K

Enable Q8/Q4 KV quantization to scale batch concurrency

Weights: 142.53 GBKV Cache (1×16K): 3.15 GBOverhead: 3.3 GB
Reserves ~10% unified memory for OS
🏃 Smooth & Snappy~9.5s for a 500-token response
Reference Launch CLI:
OLLAMA_CONTEXT_LENGTH=16384 ollama run qwen3:235b-a22b

Mac Studio M5 Ultra 512G Fits Qwen3-235B-A22B (MoE) · Cost & TCO

Approx. $9,499 (~$19/GB VRAM) · 190W TDP

📊 Qwen3-235B-A22B (MoE) Quantization VRAM Breakdown (16K Context)

Quantization Bits/Weight Weights KV Cache (16K) Total VRAM Quality Note
FP16 / BF16 16 bpw 470.2 GB 3.15 GB 483.21 GB Lossless Full Precision
FP8 / Q8_0 8.5 bpw 249.79 GB 3.15 GB 258.39 GB 99.9% Near-Lossless
GGUF Q6_K 6.56 bpw 192.78 GB 3.15 GB 200.24 GB Ultra High Quality
GGUF Q5_K_M 5.69 bpw 167.21 GB 3.15 GB 174.16 GB High Balanced Quality
Q4_K_M / AWQ 4.85 bpw 142.53 GB 3.15 GB 148.98 GB Sweet Spot (Recommended)
GGUF IQ4_XS 4.25 bpw 124.9 GB 3.15 GB 131 GB Compact 4-bit
GGUF Q3_K_M 3.91 bpw 114.91 GB 3.15 GB 120.81 GB Moderate Perplexity Loss
GGUF Q2_K 3.15 bpw 92.57 GB 3.15 GB 98.03 GB Extreme Compression

🖥️ GPU & Apple Silicon Compatibility for Qwen3-235B-A22B (MoE) (Q4_K_M)

Frequently Asked Questions

What are the exact architecture parameters of Qwen3-235B-A22B (MoE)?

Qwen3-235B-A22B (MoE) uses Qwen3MoeForCausalLM with 94 layers (94-Layer Flagship MoE (128 Experts, 8 Activated)), hidden_size=4096, 64 Q heads / 4 KV heads (head_dim=128), and vocab_size=151,936.

How much VRAM does Qwen3-235B-A22B (MoE) require?

At 4-bit quantization (Q4_K_M, 16K context), Qwen3-235B-A22B (MoE) requires 148.98 GB VRAM. Full FP16 requires 483.21 GB.