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
DeepSeek
Open-Weight DeepSeek · Dense 32.8B · Text · BF16

DeepSeek-R1-Distill-Qwen-32B Architecture Parameters, Benchmarks & VRAM

Widely deployed 32B distilled reasoning model for 24GB GPUs.

🧬 DeepSeek-R1-Distill-Qwen-32B Architecture & Tensor Specifications

Verified directly against HuggingFace (deepseek-ai/DeepSeek-R1-Distill-Qwen-32B) config.json & Model Card

HF Class: Qwen2ForCausalLM

1. Backbone & Dimensions

qwen2
Total / Active Params
32.8B (Dense)
Hidden Size (d_model)
5,120
FFN Intermediate Size
27,648
Activation / Norm
SiLU (SwiGLU) · ε=1e-5
Native Tensor Dtype
BF16

2. Attention & Layer Layout

GQA
Num Hidden Layers
64 Layers
Hidden Layer Layout
64-Layer Qwen2.5-32B Dense Backbone (Distilled from DeepSeek-R1)
Attention Heads (Q / KV)
40 Q / 8 KV (5:1 GQA)
Head Dimension
128

3. Context, RoPE & Tokenizer

Tokenizer
Native Context Length
131,072 tokens
RoPE Configuration
RoPE (θ=1,000,000 · 128K Context)
Vocab Size
152,064 (Untied)
Special Token IDs
BOS: 151646 · EOS: 151643

📋 DeepSeek-R1-Distill-Qwen-32B Official API & System Card Specifications

Verified against official DeepSeek developer documentation & pricing

API Model ID: deepseek-r1-distill-qwen-32b

1. Identity & Capabilities

API Spec
API Model ID
deepseek-r1-distill-qwen-32b
Release Date
2025-01-20
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
16,000 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
¥ 2
Cached Input / 1M
¥ 0.4 -80%
Output / 1M Tokens
¥ 6
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 ~1428 AA Index 40
🧠 Reasoning
AIME 2024 72.6 %

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

🧠 Reasoning
MATH-500 94.3 %

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
GPQA Diamond 62.1 %

PhD-level Google-proof science reasoning questions

💻 Coding
LiveCodeBench 57.2 %

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

📚 Knowledge
MMLU-Redux 88.2 %

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
C-Eval 82.2 %

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

🎛️ Official Sampling & Reasoning Parameters

Recommended generation settings & API parameter constraints

Reasoning Recommended Config
temperature=0.6 top_p=0.95 max_tokens=32,768

💡 Keep temperature in 0.5–0.7 (0.6 best); avoid system prompts and place all instructions in user message.

📊 DeepSeek-R1-Distill-Qwen-32B Full Benchmark Scorecard

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

⚔️ Compare head-to-head

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

⚡ DeepSeek-R1-Distill-Qwen-32B Interactive VRAM & Token Speed Calculator

Concurrency (Batch):
Dense 32.8B·64L (d=5120)·40Q/8KV·Ctx 128K
Runnable Models →
GPUs
32 GB VRAM·1792 GB/s·Safe Usable 31.4 GB
Smooth Fit25.03 / 31.4 GB~58.6 tok/s
3. Weight Quantization
4. Per-Request Context16K (16,384 tok)
Verified: 2026-11-20·Roofline bandwidth/compute estimate (Batch=1)
Smooth Fit
25.03/ 31.4 GB usable

Physical 32 GB · Load 79.8%

Theoretical Decode Speed
~58.6tokens/s

Actual speed varies ~±25% by runtime

Max Context at Batch=1
39K/ 128K

Enable Q8/Q4 KV quantization to scale batch concurrency

Weights: 19.88 GBKV Cache (1×16K): 4.29 GBOverhead: 0.85 GB
🚀 Blazing Fast (Instant CoT)~8.5s for a 500-token response
Reference Launch CLI:
OLLAMA_CONTEXT_LENGTH=16384 ollama run deepseek-r1:32b

RTX 5090 Fits DeepSeek-R1-Distill-Qwen-32B · Cost & TCO

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

📊 DeepSeek-R1-Distill-Qwen-32B Quantization VRAM Breakdown (16K Context)

Quantization Bits/Weight Weights KV Cache (16K) Total VRAM Quality Note
FP16 / BF16 16 bpw 65.6 GB 4.29 GB 71.66 GB Lossless Full Precision
FP8 / Q8_0 8.5 bpw 34.85 GB 4.29 GB 40.29 GB 99.9% Near-Lossless
GGUF Q6_K 6.56 bpw 26.9 GB 4.29 GB 32.18 GB Ultra High Quality
GGUF Q5_K_M 5.69 bpw 23.33 GB 4.29 GB 28.54 GB High Balanced Quality
Q4_K_M / AWQ 4.85 bpw 19.88 GB 4.29 GB 25.03 GB Sweet Spot (Recommended)
GGUF IQ4_XS 4.25 bpw 17.42 GB 4.29 GB 22.52 GB Compact 4-bit
GGUF Q3_K_M 3.91 bpw 16.03 GB 4.29 GB 21.1 GB Moderate Perplexity Loss
GGUF Q2_K 3.15 bpw 12.91 GB 4.29 GB 17.92 GB Extreme Compression

🖥️ GPU & Apple Silicon Compatibility for DeepSeek-R1-Distill-Qwen-32B (Q4_K_M)

Frequently Asked Questions

What are the exact architecture parameters of DeepSeek-R1-Distill-Qwen-32B?

DeepSeek-R1-Distill-Qwen-32B uses Qwen2ForCausalLM with 64 layers (64-Layer Qwen2.5-32B Dense Backbone (Distilled from DeepSeek-R1)), hidden_size=5120, 40 Q heads / 8 KV heads (head_dim=128), and vocab_size=152,064.

How much VRAM does DeepSeek-R1-Distill-Qwen-32B require?

At 4-bit quantization (Q4_K_M, 16K context), DeepSeek-R1-Distill-Qwen-32B requires 25.03 GB VRAM. Full FP16 requires 71.66 GB.