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
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
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
⚡ Qwen3.8-2.4T-A95B (MoE) Interactive VRAM & Token Speed Calculator
Physical 96 GB · Load 1610.5%
Exceeds usable VRAM by 1421.07 GB
Enable Q8/Q4 KV quantization to scale batch concurrency
OLLAMA_CONTEXT_LENGTH=16384 ollama run qwen3-8-2-4t-a95bInsufficient 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.