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 2.45T / A95B · Text · BF16 / FP8

Qwen3.8-2.4T-A95B (MoE) Architecture Parameters, Benchmarks & VRAM

2.44T total / 95B active flagship Qwen3.8 MoE backbone (92 layers, 512 experts).

🧬 Qwen3.8-2.4T-A95B (MoE) Architecture & Tensor Specifications

Verified directly against HuggingFace (Qwen/Qwen3.8-2.4T-A95B) config.json & Model Card

HF Class: Qwen3_5MoeForCausalLM

1. Backbone & Dimensions

qwen3_5_moe_text
Total / Active Params
2.446T / 95B
Hidden Size (d_model)
8,192
MoE Expert Dim
2,048
MoE Experts
512 Routed · Top-10
Activation / Norm
SiLU (SwiGLU) · ε=0.000001
Native Tensor Dtype
BF16 / FP8

2. Attention & Layer Layout

HYBRID
Num Hidden Layers
92 Layers (23 Global KV)
Hidden Layer Layout
92-Layer Sparse MoE (512 Experts, Top-10 Active)
Attention Heads (Q / KV)
64 Q / 4 KV (16:1 GQA)
Head Dimension
256

3. Context, RoPE & Tokenizer

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

📋 Qwen3.8-2.4T-A95B (MoE) Official API & System Card Specifications

Verified against official Alibaba Qwen developer documentation & pricing

API Model ID: qwen3.8-max

1. Identity & Capabilities

API Spec
API Model ID
qwen3.8-max
Release Date
2026-08-03
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
131,000 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
¥ 12
Cached Input / 1M
¥ 1.5 -88%
Output / 1M Tokens
¥ 36
Tiered Pricing Notes
Alibaba Bailian: Hosted on Alibaba Cloud Bailian (CN Direct)

🎛️ Official Sampling & Reasoning Parameters

Recommended generation settings & API parameter constraints

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

⚡ Qwen3.8-2.4T-A95B (MoE) Interactive VRAM & Token Speed Calculator

Concurrency (Batch):
MoE 2446B / A95B·92L (d=8192)·64Q/4KV·Ctx 256K
Runnable Models →
GPUs
96 GB VRAM·4000 GB/s·Safe Usable 94.1 GB
Out of VRAM (OOM)1515.15 / 94.1 GBOOM
3. Weight Quantization
4. Per-Request Context16K (16,384 tok)
Verified: 2026-11-20·Roofline bandwidth/compute estimate (Batch=1)
Out of VRAM (OOM)
1515.15/ 94.1 GB usable

Physical 96 GB · Load 1610.5%

Theoretical Decode Speed
—tokens/s

Exceeds usable VRAM by 1421.07 GB

Max Context at Batch=1
0K/ 256K

Enable Q8/Q4 KV quantization to scale batch concurrency

Weights: 1482.89 GBKV Cache (1×16K): 2.15 GBOverhead: 30.11 GB
⛔ Cannot Load
Reference Launch CLI:
OLLAMA_CONTEXT_LENGTH=16384 ollama run qwen3-8-2-4t-a95b

Insufficient Local VRAM? Ways to Run Qwen3.8-2.4T-A95B (MoE)

Rent an hourly high-VRAM cloud GPU pod, or check the 12-month Buy vs Cloud vs API break-even curve.

📊 Qwen3.8-2.4T-A95B (MoE) Quantization VRAM Breakdown (16K Context)

Quantization Bits/Weight Weights KV Cache (16K) Total VRAM Quality Note
FP16 / BF16 16 bpw 4892 GB 2.15 GB 4992.44 GB Lossless Full Precision
FP8 / Q8_0 8.5 bpw 2598.88 GB 2.15 GB 2653.45 GB 99.9% Near-Lossless
GGUF Q6_K 6.56 bpw 2005.72 GB 2.15 GB 2048.44 GB Ultra High Quality
GGUF Q5_K_M 5.69 bpw 1739.72 GB 2.15 GB 1777.11 GB High Balanced Quality
Q4_K_M / AWQ 4.85 bpw 1482.89 GB 2.15 GB 1515.15 GB Sweet Spot (Recommended)
GGUF IQ4_XS 4.25 bpw 1299.44 GB 2.15 GB 1328.03 GB Compact 4-bit
GGUF Q3_K_M 3.91 bpw 1195.48 GB 2.15 GB 1221.99 GB Moderate Perplexity Loss
GGUF Q2_K 3.15 bpw 963.11 GB 2.15 GB 984.98 GB Extreme Compression

🖥️ GPU & Apple Silicon Compatibility for Qwen3.8-2.4T-A95B (MoE) (Q4_K_M)

Frequently Asked Questions

What are the exact architecture parameters of Qwen3.8-2.4T-A95B (MoE)?

Qwen3.8-2.4T-A95B (MoE) uses Qwen3_5MoeForCausalLM with 92 layers (92-Layer Sparse MoE (512 Experts, Top-10 Active)), hidden_size=8192, 64 Q heads / 4 KV heads (head_dim=256), and vocab_size=248,320.

How much VRAM does Qwen3.8-2.4T-A95B (MoE) require?

At 4-bit quantization (Q4_K_M, 16K context), Qwen3.8-2.4T-A95B (MoE) requires 1515.15 GB VRAM. Full FP16 requires 4992.44 GB.