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 · Dense 8.2B · Text · BF16

Qwen3-8B Architecture Parameters, Benchmarks & VRAM

Fast 8.2B model running smoothly on 8GB GPUs and 16GB unified memory.

🧬 Qwen3-8B Architecture & Tensor Specifications

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

HF Class: Qwen3ForCausalLM

1. Backbone & Dimensions

qwen3
Total / Active Params
8.2B (Dense)
Non-Embedding Params
6.95B
Hidden Size (d_model)
4,096
FFN Intermediate Size
12,288
Activation / Norm
SiLU (SwiGLU) · ε=1e-6
Native Tensor Dtype
BF16

2. Attention & Layer Layout

GQA
Num Hidden Layers
36 Layers
Hidden Layer Layout
36-Layer Compact Dense Transformer (GQA + SwiGLU + QK-Norm)
Attention Heads (Q / KV)
32 Q / 8 KV (4:1 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

📋 Qwen3-8B Official API & System Card Specifications

Verified against official Alibaba Qwen developer documentation & pricing

API Model ID: Qwen/Qwen3-8B

1. Identity & Capabilities

API Spec
API Model ID
Qwen/Qwen3-8B
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
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
¥ 0
Cached Input / 1M
¥ 0
Output / 1M Tokens
¥ 0
Tiered Pricing Notes
SiliconFlow: 100% Free tier (rate-limited)

🏆 Official Benchmark Results

Sourced from official System Card / Model Card & technical reports

Arena Elo ~1348 AA Index 29
📚 Knowledge Thinking
MMLU-Redux 87.5 %

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.

📚 Knowledge Thinking
C-Eval 83.4 %

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

🧠 Reasoning Thinking
GPQA Diamond 62 %

PhD-level Google-proof science reasoning questions

🧠 Reasoning Thinking
MATH-500 97.4 %

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 76 %

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

🧠 Reasoning Thinking
AIME 2025 67.3 %

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 57.5 %

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

🤖 Agentic Thinking
BFCL v3 68.1 %

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.

⚡ Qwen3-8B Interactive VRAM & Token Speed Calculator

Concurrency (Batch):
Dense 8.2B·36L (d=4096)·32Q/8KV·Ctx 128K
Runnable Models →
GPUs
32 GB VRAM·1792 GB/s·Safe Usable 31.4 GB
Smooth Fit7.94 / 31.4 GB~208.8 tok/s
3. Weight Quantization
4. Per-Request Context16K (16,384 tok)
Verified: 2026-11-20·Roofline bandwidth/compute estimate (Batch=1)
Smooth Fit
7.94/ 31.4 GB usable

Physical 32 GB · Load 25.3%

Theoretical Decode Speed
~208.8tokens/s

Actual speed varies ~±25% by runtime

Max Context at Batch=1
128K/ 128K

Enable Q8/Q4 KV quantization to scale batch concurrency

Weights: 4.97 GBKV Cache (1×16K): 2.42 GBOverhead: 0.55 GB
🚀 Blazing Fast (Instant CoT)~2.4s for a 500-token response
Reference Launch CLI:
OLLAMA_CONTEXT_LENGTH=16384 ollama run qwen3:8b

RTX 5090 Fits Qwen3-8B · Cost & TCO

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

📊 Qwen3-8B Quantization VRAM Breakdown (16K Context)

Quantization Bits/Weight Weights KV Cache (16K) Total VRAM Quality Note
FP16 / BF16 16 bpw 16.4 GB 2.42 GB 19.59 GB Lossless Full Precision
FP8 / Q8_0 8.5 bpw 8.71 GB 2.42 GB 11.75 GB 99.9% Near-Lossless
GGUF Q6_K 6.56 bpw 6.72 GB 2.42 GB 9.72 GB Ultra High Quality
GGUF Q5_K_M 5.69 bpw 5.83 GB 2.42 GB 8.81 GB High Balanced Quality
Q4_K_M / AWQ 4.85 bpw 4.97 GB 2.42 GB 7.94 GB Sweet Spot (Recommended)
GGUF IQ4_XS 4.25 bpw 4.36 GB 2.42 GB 7.31 GB Compact 4-bit
GGUF Q3_K_M 3.91 bpw 4.01 GB 2.42 GB 6.95 GB Moderate Perplexity Loss
GGUF Q2_K 3.15 bpw 3.23 GB 2.42 GB 6.16 GB Extreme Compression

🖥️ GPU & Apple Silicon Compatibility for Qwen3-8B (Q4_K_M)

Frequently Asked Questions

What are the exact architecture parameters of Qwen3-8B?

Qwen3-8B uses Qwen3ForCausalLM with 36 layers (36-Layer Compact Dense Transformer (GQA + SwiGLU + QK-Norm)), hidden_size=4096, 32 Q heads / 8 KV heads (head_dim=128), and vocab_size=151,936.

How much VRAM does Qwen3-8B require?

At 4-bit quantization (Q4_K_M, 16K context), Qwen3-8B requires 7.94 GB VRAM. Full FP16 requires 19.59 GB.