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 397B / A17B · Text+Image+Video · BF16

Qwen3.5-397B-A17B (MoE) Architecture Parameters, Benchmarks & VRAM

Largest Qwen3.5 open-weight flagship (397B total / 17B active, 512 experts) with native vision & video.

🧬 Qwen3.5-397B-A17B (MoE) Architecture & Tensor Specifications

Verified directly against HuggingFace (Qwen/Qwen3.5-397B-A17B) config.json & Model Card

HF Class: Qwen3_5MoeForConditionalGeneration

1. Backbone & Dimensions

qwen3_5_moe
Total / Active Params
397B / 17B
Hidden Size (d_model)
4,096
MoE Expert Dim
1,024
MoE Experts
512 Routed · Top-10 + 1 Shared
Activation / Norm
SiLU (SwiGLU) · ε=1e-6
Native Tensor Dtype
BF16

2. Attention & Layer Layout

HYBRID
Num Hidden Layers
60 Layers (15 Global KV)
Hidden Layer Layout
60 Hybrid MoE Layers: 15 × [3 × Gated DeltaNet→MoE + 1 × Gated Attention→MoE]
Attention Heads (Q / KV)
32 Q / 2 KV (16:1 GQA)
Head Dimension
256
Linear / KV Compression
Only 15 layers × 2 KV heads allocate KV cache; extensible to 1M tokens via YaRN
Vision Encoder / MTP
ViT 27L · MTP 1L

3. Context, RoPE & Tokenizer

Tokenizer
Native Context Length
262,144 tokens
Max Extended Context
1,010,000 tokens
RoPE Configuration
mRoPE (θ=10,000,000 · partial=0.25 · YaRN 4x→1M)
Vocab Size
248,320 (Untied)
Special Token IDs
BOS: 248044 · EOS: [248046,248044] · PAD: 248044

📋 Qwen3.5-397B-A17B (MoE) Official API & System Card Specifications

Verified against official Alibaba Qwen developer documentation & pricing

API Model ID: qwen3.5-397b-a17b

1. Identity & Capabilities

API Spec
API Model ID
qwen3.5-397b-a17b
Release Date
2026-02-23
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
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
¥ 1.2
Cached Input / 1M
¥ 0.24 -80%
Output / 1M Tokens
¥ 7.19
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 ~1495 AA Index 51
📚 Knowledge
MMLU-Pro 87.8 %

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
C-Eval 93 %

C-Eval is a comprehensive Chinese evaluation suite designed to assess advanced knowledge and reasoning abilities of foundation models in a Chinese context.

🧠 Reasoning
GPQA Diamond 88.4 %

PhD-level Google-proof science reasoning questions

🧠 Reasoning
AIME 2026 91.3 %

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.

💻 Coding
LiveCodeBench v6 83.6 %

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

🤖 Agentic
TAU2-Bench 86.7 %

Dual-control conversational tool-use benchmark

🤖 Agentic
BFCL-V4 72.9 %

Berkeley Function Calling Leaderboard V4 (BFCL-V4) evaluates LLMs on their ability to accurately call functions and APIs, including simple, multiple, parallel, and nested function calls across diverse programming scenarios.

🎛️ Official Sampling & Reasoning Parameters

Recommended generation settings & API parameter constraints

Thinking Mode
temperature=0.6 top_p=0.95 top_k=20 min_p=0 max_tokens=32,768
Instruct Mode
temperature=0.7 top_p=0.8 top_k=20 presence_penalty=1.5 max_tokens=32,768

📊 Qwen3.5-397B-A17B (MoE) Full Benchmark Scorecard

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

⚔️ Compare head-to-head

🤖 Agents & Tools 10 items

t2-bench #4/12 86.7%

t2-bench is a benchmark for evaluating agentic tool use capabilities, measuring how well models can select, sequence, and utilize tools to solve complex tasks. It tests autonomous planning and execution in multi-step scenarios.

WideSearch #6/13 74%

WideSearch is an agentic search benchmark that evaluates models' ability to perform broad, parallel search operations across multiple sources. It tests wide-coverage information retrieval and synthesis capabilities.

BFCL-V4 #3/10 72.9%

Berkeley Function Calling Leaderboard V4 (BFCL-V4) evaluates LLMs on their ability to accurately call functions and APIs, including simple, multiple, parallel, and nested function calls across diverse programming scenarios.

BrowseComp-zh #2/7 70.3%

A high-difficulty benchmark purpose-built to comprehensively evaluate LLM agents on the Chinese web, consisting of 289 multi-hop questions spanning 11 diverse domains including Film & TV, Technology, Medicine, and History.

BrowseComp #35/54 69%

BrowseComp is a benchmark comprising 1,266 questions that challenge AI agents to persistently navigate the internet in search of hard-to-find, entangled information.

VITA-Bench #1/8 49.7%

VITA-Bench evaluates AI agents on real-world virtual task automation, measuring their ability to complete complex multi-step tasks in simulated environments.

Seal-0 #3/5 46.9%

Seal-0 is a benchmark for evaluating agentic search capabilities, testing models' ability to navigate and retrieve information using tools.

MCP-Mark #6/7 46.1%

MCP-Mark evaluates LLMs on their ability to use Model Context Protocol (MCP) tools effectively, testing tool discovery, selection, invocation, and result interpretation across diverse MCP server scenarios.

Toolathlon #29/32 38.3%

Tool Decathlon is a comprehensive benchmark for evaluating AI agents' ability to use multiple tools across diverse task categories. It measures proficiency in tool selection, sequencing, and execution across ten different tool-use scenarios.

DeepPlanning #3/7 34.3%

DeepPlanning evaluates LLMs on complex multi-step planning tasks requiring long-horizon reasoning, goal decomposition, and strategic decision-making.

📐 Math 11 items

MMLU-Redux #2/15 94.9%

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.

HMMT 2025 #3/11 94.8%

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 #2/9 92.7%

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.

AIME 2026 #13/18 91.3%

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.

MMMLU #10/20 88.5%

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…

MMLU-Pro #5/39 87.8%

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.

MMLU-ProX #3/13 84.7%

Extended version of MMLU-Pro providing additional challenging multiple-choice questions for evaluating language models across diverse academic and professional domains. Built on the foundation of the Massive Multitask Language Understanding benchmark framework.

IMO-AnswerBench #15/17 80.9%

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

PolyMATH #4/9 73.3%

Polymath is a challenging multi-modal mathematical reasoning benchmark designed to evaluate the general cognitive reasoning abilities of Multi-modal Large Language Models (MLLMs).

SuperGPQA #5/16 70.4%

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

Humanity's Last Exam #45/69 28.7%

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.

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

⚡ Qwen3.5-397B-A17B (MoE) Interactive VRAM & Token Speed Calculator

Concurrency (Batch):
MoE 397B / A17B·60L (d=4096)·32Q/2KV·Ctx 256K·Text+Image+Video
Runnable Models →
512 GB Unified·1200 GB/s·Safe Usable 460.8 GB
Smooth Fit247.5 / 460.8 GB~74.3 tok/s
3. Weight Quantization
4. Per-Request Context16K (16,384 tok)
Verified: 2026-11-20·Roofline bandwidth/compute estimate (Batch=1)
Smooth Fit
247.5/ 460.8 GB usable

Physical 512 GB · Load 53.7%

Theoretical Decode Speed
~74.3tokens/s

MoE: reads only 10.31 GB active weights/step

Max Context at Batch=1
256K/ 256K

Enable Q8/Q4 KV quantization to scale batch concurrency

Weights: 240.68 GBKV Cache (1×16K): 0.7 GBOverhead: 6.11 GB
Reserves ~10% unified memory for OS
🚀 Blazing Fast (Instant CoT)~6.7s for a 500-token response
Reference Launch CLI:
OLLAMA_CONTEXT_LENGTH=16384 ollama run qwen3.5:397b-a17b

Mac Studio M5 Ultra 512G Fits Qwen3.5-397B-A17B (MoE) · Cost & TCO

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

📊 Qwen3.5-397B-A17B (MoE) Quantization VRAM Breakdown (16K Context)

Quantization Bits/Weight Weights KV Cache (16K) Total VRAM Quality Note
FP16 / BF16 16 bpw 794 GB 0.7 GB 811.88 GB Lossless Full Precision
FP8 / Q8_0 8.5 bpw 421.81 GB 0.7 GB 432.25 GB 99.9% Near-Lossless
GGUF Q6_K 6.56 bpw 325.54 GB 0.7 GB 334.05 GB Ultra High Quality
GGUF Q5_K_M 5.69 bpw 282.37 GB 0.7 GB 290.02 GB High Balanced Quality
Q4_K_M / AWQ 4.85 bpw 240.68 GB 0.7 GB 247.5 GB Sweet Spot (Recommended)
GGUF IQ4_XS 4.25 bpw 210.91 GB 0.7 GB 217.13 GB Compact 4-bit
GGUF Q3_K_M 3.91 bpw 194.03 GB 0.7 GB 199.92 GB Moderate Perplexity Loss
GGUF Q2_K 3.15 bpw 156.32 GB 0.7 GB 161.45 GB Extreme Compression

🖥️ GPU & Apple Silicon Compatibility for Qwen3.5-397B-A17B (MoE) (Q4_K_M)

Frequently Asked Questions

What are the exact architecture parameters of Qwen3.5-397B-A17B (MoE)?

Qwen3.5-397B-A17B (MoE) uses Qwen3_5MoeForConditionalGeneration with 60 layers (60 Hybrid MoE Layers: 15 × [3 × Gated DeltaNet→MoE + 1 × Gated Attention→MoE]), hidden_size=4096, 32 Q heads / 2 KV heads (head_dim=256), and vocab_size=248,320.

How much VRAM does Qwen3.5-397B-A17B (MoE) require?

At 4-bit quantization (Q4_K_M, 16K context), Qwen3.5-397B-A17B (MoE) requires 247.5 GB VRAM. Full FP16 requires 811.88 GB.