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 36B / A3B · Text+Image+Video · BF16 / FP8

Qwen3.6-35B-A3B (MoE) Architecture Parameters, Benchmarks & VRAM

Qwen3.6 36B total / 3B active multimodal MoE tuned for coding stability and fast local inference.

🧬 Qwen3.6-35B-A3B (MoE) Architecture & Tensor Specifications

Verified directly against HuggingFace (Qwen/Qwen3.6-35B-A3B) config.json & Model Card

HF Class: Qwen3_5MoeForConditionalGeneration

1. Backbone & Dimensions

qwen3_5_moe
Total / Active Params
36B / 3B
Hidden Size (d_model)
2,048
MoE Expert Dim
512
MoE Experts
256 Routed · Top-8
Activation / Norm
SiLU (SwiGLU) · ε=0.000001
Native Tensor Dtype
BF16 / FP8

2. Attention & Layer Layout

HYBRID
Num Hidden Layers
40 Layers (10 Global KV)
Hidden Layer Layout
40-Layer Sparse MoE (256 Experts, Top-8 Active)
Attention Heads (Q / KV)
16 Q / 2 KV (8:1 GQA)
Head Dimension
256
Vision Encoder / MTP
ViT 27L

3. Context, RoPE & Tokenizer

Tokenizer
Native Context Length
262,144 tokens
RoPE Configuration
RoPE (原生 262,144 Context)
Vocab Size
248,320 (Untied)

📋 Qwen3.6-35B-A3B (MoE) Official API & System Card Specifications

Verified against official Alibaba Qwen developer documentation & pricing

API Model ID: Qwen/Qwen3.6-35B-A3B

1. Identity & Capabilities

API Spec
API Model ID
Qwen/Qwen3.6-35B-A3B
Release Date
2026-04-17
Weight License
Open-Weight (API Spec)
Native Reasoning
—
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
256,000 tokens (256K)
Max Output Tokens
66,000 tokens
Input Modalities
Text + Image + Video
Output Modalities
Text
Direct China Access
✅ Direct CN + RMB
Deployment Mode
Open-Weight + Cloud API

3. Official Pricing & Cache

CNY / 1M
Input / 1M Tokens
¥ 1.8
Cached Input / 1M
¥ 1.8
Output / 1M Tokens
¥ 10.76
Tiered Pricing Notes
SiliconFlow: Open-weight cloud hosting on SiliconFlow

🎛️ Official Sampling & Reasoning Parameters

Recommended generation settings & API parameter constraints

Official Default Sampling
temperature=0.7 top_p=0.95 max_tokens=32,768

📊 Qwen3.6-35B-A3B (MoE) Full Benchmark Scorecard

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

⚔️ Compare head-to-head

💻 Coding 9 items

SWE-Bench Verified #32/59 73.4%

A verified subset of 500 software engineering problems from real GitHub issues, validated by human annotators for evaluating language models' ability to resolve real-world coding issues by generating patches for Python codebases.

SWE-bench Multilingual #26/33 67.2%

A multilingual benchmark for issue resolving in software engineering that covers Java, TypeScript, JavaScript, Go, Rust, C, and C++.

MCP Atlas #30/39 62.8%

MCP Atlas is a benchmark for evaluating AI models on scaled tool use capabilities, measuring how well models can coordinate and utilize multiple tools across complex multi-step tasks.

ZClawBench #4/4 52.6%

ZClawBench evaluates Claw-style agent task execution quality, measuring a model's ability to autonomously complete complex multi-step coding tasks in real-world environments.

Terminal-Bench 2.0 #25/38 51.5%

Terminal-Bench 2.0 is an updated benchmark for testing AI agents' tool use ability to operate a computer via terminal.

Claw-Eval #13/13 50%

Claw-Eval tests real-world agentic task completion across complex multi-step scenarios, evaluating a model's ability to use tools, navigate environments, and complete end-to-end tasks autonomously.

SWE-Bench Pro #49/52 49.5%

SWE-Bench Pro is an advanced version of SWE-Bench that evaluates language models on complex, real-world software engineering tasks requiring extended reasoning and multi-step problem solving.

NL2Repo #22/23 29.4%

NL2Repo evaluates long-horizon coding capabilities including repository-level understanding, where models must generate or modify code across entire repositories from natural language specifications.

SkillsBench #10/11 28.7%

SkillsBench evaluates coding agents on self-contained programming tasks, measuring practical engineering skills across diverse software development scenarios.

📐 Math 10 items

MMLU-Redux #7/15 93.3%

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.

AIME 2026 #12/18 92.7%

All 30 problems from the 2026 American Invitational Mathematics Examination (AIME I and AIME II), testing olympiad-level mathematical reasoning with integer answers from 000-999.

HMMT 2025 #10/11 90.7%

Harvard-MIT Mathematics Tournament 2025 - A prestigious student-organized mathematics competition for high school students featuring two tournaments (November 2025 at MIT and February 2026 at Harvard) with individual tests, team rounds, and guts rounds.

HMMT25 #7/9 89.1%

Harvard-MIT Mathematics Tournament 2025 - A prestigious student-organized mathematics competition for high school students featuring two tournaments (November 2025 at MIT and February 2026 at Harvard) with individual tests, team rounds, and guts rounds.

MathVista-Mini #5/7 86.4%

MathVista-Mini is a smaller version of the MathVista benchmark that evaluates mathematical reasoning in visual contexts.

MMLU-Pro #15/39 85.2%

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.

HMMT Feb 26 #10/11 83.6%

HMMT February 2026 is a math competition benchmark based on problems from the Harvard-MIT Mathematics Tournament, testing advanced mathematical problem-solving and reasoning.

IMO-AnswerBench #17/17 78.9%

IMO-AnswerBench is a benchmark for evaluating mathematical reasoning capabilities on International Mathematical Olympiad (IMO) problems, focusing on answer generation and verification.

SuperGPQA #11/16 64.7%

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

Humanity's Last Exam #53/69 21.4%

Humanity's Last Exam (HLE) is a multi-modal academic benchmark with 2,500 questions across mathematics, humanities, and natural sciences, designed to test LLM capabilities at the frontier of human knowledge with unambiguous, verifiable solutions.

👁️ Multimodal 21 items

MMBench-V1.1 #1/6 92.8%

Version 1.1 of MMBench, an improved bilingual benchmark for assessing multi-modal capabilities of vision-language models through multiple-choice questions in both English and Chinese, providing systematic evaluation across diverse vision-language tasks.

AI2D #4/7 92.7%

AI2D is a dataset of 4,903 illustrative diagrams from grade school natural sciences (such as food webs, human physiology, and life cycles) with over 15,000 multiple choice questions and answers.

OmniDocBench 1.5 #5/16 89.9%

OmniDocBench 1.5 is a comprehensive benchmark for evaluating multimodal large language models on document understanding tasks, including OCR, document parsing, information extraction, and visual question answering across diverse document types.

VideoMME w sub. #5/8 86.6%

The first-ever comprehensive evaluation benchmark of Multi-modal LLMs in Video analysis. Features 900 videos (254 hours) with 2,700 question-answer pairs covering 6 primary visual domains and 30 subfields.

MLVU #5/7 86.2%

A comprehensive benchmark for multi-task long video understanding that evaluates multimodal large language models on videos ranging from 3 minutes to 2 hours across 9 distinct tasks including reasoning, captioning, recognition, and summarization.

RealWorldQA #7/13 85.3%

RealWorldQA is a benchmark designed to evaluate basic real-world spatial understanding capabilities of multimodal models.

VideoMMMU #8/14 83.7%

Video-MMMU evaluates Large Multimodal Models' ability to acquire knowledge from expert-level professional videos across six disciplines through three cognitive stages: perception, comprehension, and adaptation.

VideoMME w/o sub. #3/4 82.5%

Video-MME is a comprehensive evaluation benchmark for multi-modal large language models in video analysis. It features 900 videos across 6 primary visual domains with 30 subfields, ranging from 11 seconds to 1 hour in duration, with 2,700 question-answer pairs.

CC-OCR #2/6 81.9%

A comprehensive OCR benchmark for evaluating Large Multimodal Models (LMMs) in literacy. Comprises four OCR-centric tracks: multi-scene text reading, multilingual text reading, document parsing, and key information extraction.

MMMU #6/16 81.7%

MMMU (Massive Multi-discipline Multimodal Understanding) is a benchmark designed to evaluate multimodal models on college-level subject knowledge and deliberate reasoning.

CharXiv-R #29/41 78%

CharXiv-R is the reasoning component of the CharXiv benchmark, focusing on complex reasoning questions that require synthesizing information across visual chart elements.

MMMU-Pro #28/44 75.3%

A more robust multi-discipline multimodal understanding benchmark that enhances MMMU through a three-step process: filtering text-only answerable questions, augmenting candidate options, and introducing vision-only input settings.

MVBench #5/7 74.6%

A comprehensive multi-modal video understanding benchmark covering 20 challenging video tasks that require temporal understanding beyond single-frame analysis.

LVBench #13/15 71.4%

LVBench is an extreme long video understanding benchmark designed to evaluate multimodal models on videos up to two hours in duration. It contains 6 major categories and 21 subcategories, with videos averaging five times longer than existing datasets.

Hallusion Bench #2/6 69.8%

A comprehensive benchmark designed to evaluate image-context reasoning in large visual-language models (LVLMs) by challenging models with 346 images and 1,129 carefully crafted questions to assess language hallucination and visual illusion.

ODinW #3/7 50.8%

Object Detection in the Wild (ODinW) benchmark for evaluating object detection models' task-level transfer ability across diverse real-world datasets in terms of prediction accuracy and adaptation efficiency.

RefCOCO-avg #3/6 0.9

RefCOCO-avg measures object grounding accuracy averaged across RefCOCO, RefCOCO+, and RefCOCOg benchmarks.

EmbSpatialBench #3/6 0.8

EmbSpatialBench evaluates embodied spatial understanding and reasoning capabilities.

RefSpatialBench #5/7 0.6

RefSpatialBench evaluates spatial reference understanding and grounding.

SimpleVQA #10/13 0.6

SimpleVQA is a visual question answering benchmark focused on simple queries.

ZEROBench-Sub #4/5 0.3

ZEROBench-Sub is a subset of the ZEROBench benchmark.

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

⚡ Qwen3.6-35B-A3B (MoE) Interactive VRAM & Token Speed Calculator

Concurrency (Batch):
MoE 36B / A3B·40L (d=2048)·16Q/2KV·Ctx 256K·Text+Image+Video
Runnable Models →
GPUs
32 GB VRAM·1792 GB/s·Safe Usable 31.4 GB
Smooth Fit23.96 / 31.4 GB~420 tok/s
3. Weight Quantization
4. Per-Request Context16K (16,384 tok)
Verified: 2026-11-20·Roofline bandwidth/compute estimate (Batch=1)
Smooth Fit
23.96/ 31.4 GB usable

Physical 32 GB · Load 76.4%

Theoretical Decode Speed
~420tokens/s

MoE: reads only 1.82 GB active weights/step

Max Context at Batch=1
256K/ 256K

Enable Q8/Q4 KV quantization to scale batch concurrency

Weights: 21.83 GBKV Cache (1×16K): 0.4 GBOverhead: 1.74 GB
🚀 Blazing Fast (Instant CoT)~1.2s for a 500-token response
Reference Launch CLI:
OLLAMA_CONTEXT_LENGTH=16384 ollama run qwen3-6-35b-a3b

RTX 5090 Fits Qwen3.6-35B-A3B (MoE) · Cost & TCO

Approx. $2,199 (~$69/GB VRAM) · 575W TDP

📊 Qwen3.6-35B-A3B (MoE) Quantization VRAM Breakdown (16K Context)

Quantization Bits/Weight Weights KV Cache (16K) Total VRAM Quality Note
FP16 / BF16 16 bpw 72 GB 0.4 GB 75.14 GB Lossless Full Precision
FP8 / Q8_0 8.5 bpw 38.25 GB 0.4 GB 40.72 GB 99.9% Near-Lossless
GGUF Q6_K 6.56 bpw 29.52 GB 0.4 GB 31.81 GB Ultra High Quality
GGUF Q5_K_M 5.69 bpw 25.61 GB 0.4 GB 27.82 GB High Balanced Quality
Q4_K_M / AWQ 4.85 bpw 21.83 GB 0.4 GB 23.96 GB Sweet Spot (Recommended)
GGUF IQ4_XS 4.25 bpw 19.13 GB 0.4 GB 21.21 GB Compact 4-bit
GGUF Q3_K_M 3.91 bpw 17.6 GB 0.4 GB 19.65 GB Moderate Perplexity Loss
GGUF Q2_K 3.15 bpw 14.18 GB 0.4 GB 16.16 GB Extreme Compression

🖥️ GPU & Apple Silicon Compatibility for Qwen3.6-35B-A3B (MoE) (Q4_K_M)

Frequently Asked Questions

What are the exact architecture parameters of Qwen3.6-35B-A3B (MoE)?

Qwen3.6-35B-A3B (MoE) uses Qwen3_5MoeForConditionalGeneration with 40 layers (40-Layer Sparse MoE (256 Experts, Top-8 Active)), hidden_size=2048, 16 Q heads / 2 KV heads (head_dim=256), and vocab_size=248,320.

How much VRAM does Qwen3.6-35B-A3B (MoE) require?

At 4-bit quantization (Q4_K_M, 16K context), Qwen3.6-35B-A3B (MoE) requires 23.96 GB VRAM. Full FP16 requires 75.14 GB.