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 · MoE 1.6T / A49B · Text · FP8 Mixed

DeepSeek-V4-Pro (1.6T MoE) Architecture Parameters, Benchmarks & VRAM

1.6T total / 49B active open-weight flagship MoE with CSA+HCA hybrid sparse attention and native 1M context.

🧬 DeepSeek-V4-Pro (1.6T MoE) Architecture & Tensor Specifications

Verified directly against HuggingFace (deepseek-ai/DeepSeek-V4-Pro) config.json & Model Card

HF Class: DeepseekV4ForCausalLM

1. Backbone & Dimensions

deepseek_v4
Total / Active Params
1.6T / 49B
Hidden Size (d_model)
7,168
MoE Expert Dim
3,072
MoE Experts
384 Routed · Top-6 + 1 Shared
Activation / Norm
SiLU (SwiGLU) · ε=1e-6
Native Tensor Dtype
FP8 Mixed

2. Attention & Layer Layout

MLA
Num Hidden Layers
61 Layers
Hidden Layer Layout
61-Layer CSA + HCA Hybrid Sparse Attention MoE + mHC (384 Routed + 1 Shared Expert)
Attention Heads (Q / KV)
128 Q / 1 KV (128:1 GQA)
Head Dimension
512
Linear / KV Compression
27% single-token FLOPs and 10% KV cache vs DeepSeek-V3.2 at 1M context
Vision Encoder / MTP
MTP 1L

3. Context, RoPE & Tokenizer

Tokenizer
Native Context Length
1,048,576 tokens
RoPE Configuration
Decoupled RoPE (θ=10,000 · qk_rope_dim=64 · 原生 1,048,576 Context)
Vocab Size
129,280 (Untied)
Special Token IDs
BOS: 0 · EOS: 1

📋 DeepSeek-V4-Pro (1.6T MoE) Official API & System Card Specifications

Verified against official DeepSeek developer documentation & pricing

API Model ID: deepseek-v4-pro

1. Identity & Capabilities

API Spec
API Model ID
deepseek-v4-pro
Release Date
2026-08-13
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
1,024,000 tokens (1024K)
Max Output Tokens
384,000 tokens
Input Modalities
Text
Output Modalities
Text
Direct China Access
✅ Direct CN + RMB
Deployment Mode
Open-Weight + Cloud API

3. Official Pricing & Cache

USD / 1M
Input / 1M Tokens
$ 1.32
Cached Input / 1M
$ 0.044 -97%
Output / 1M Tokens
$ 3.96
Tiered Pricing Notes
DeepSeek Official: Standard peak rate; 50% off during off-peak ($0.66/$1.98)

🏆 Official Benchmark Results

Sourced from official System Card / Model Card & technical reports

Arena Elo ~1504 AA Index 52
📚 Knowledge Max
MMLU-Pro 87.5 %

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 Max
Chinese-SimpleQA 84.4 %

Chinese short-form factual accuracy benchmark

🧠 Reasoning Max
GPQA Diamond 90.1 %

PhD-level Google-proof science reasoning questions

🧠 Reasoning Max
HMMT 2026 Feb 95.2 %
💻 Coding Max
LiveCodeBench 93.5 %

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

💻 Coding Max
SWE-bench Verified 80.6 %

A verified subset of 500 software engineering problems from real GitHub issues, validated by human annotators for evaluating language models' ability to resolve real-world coding issues by generating patches for Python codebases.

🤖 Agentic Max
Terminal Bench 2.0 67.9 %
🤖 Agentic Max
BrowseComp 83.4 %

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.

🎛️ Official Sampling & Reasoning Parameters

Recommended generation settings & API parameter constraints

Think Max Reasoning Mode
temperature=1 top_p=1 max_tokens=65,536

💡 Set temperature=1.0, top_p=1.0; reserve >=384K context for Think Max.

📊 DeepSeek-V4-Pro (1.6T MoE) Full Benchmark Scorecard

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

⚔️ Compare head-to-head

💻 Coding 8 items

LiveCodeBench #1/18 93.5%

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

SWE-Bench Verified #7/59 80.6%

A verified subset of 500 software engineering problems from real GitHub issues, validated by human annotators for evaluating language models' ability to resolve real-world coding issues by generating patches for Python codebases.

SWE-bench Multilingual #14/33 76.2%

A multilingual benchmark for issue resolving in software engineering that covers Java, TypeScript, JavaScript, Go, Rust, C, and C++.

MCP Atlas #22/39 73.6%

MCP Atlas is a benchmark for evaluating AI models on scaled tool use capabilities, measuring how well models can coordinate and utilize multiple tools across complex multi-step tasks.

Terminal-Bench 2.0 #13/38 67.9%

Terminal-Bench 2.0 is an updated benchmark for testing AI agents' tool use ability to operate a computer via terminal.

SWE-Bench Pro #37/52 55.4%

SWE-Bench Pro is an advanced version of SWE-Bench that evaluates language models on complex, real-world software engineering tasks requiring extended reasoning and multi-step problem solving.

FrontierSWE #12/16 29%

FrontierSWE measures whether an agent can complete open-ended technical projects at the scale of hours to tens of hours, spanning systems optimization, large-scale code construction, and applied ML research. Performance is reported as a dominance score, where higher is better.

FrontierCode 1.1 #21/23 17.6%

FrontierCode 1.1 evaluates whether coding-agent changes are mergeable, using unit tests, maintainer-defined rubrics, and verifiers. Runs flagged for unfair internet use receive a zero score.

📐 Math 7 items

HMMT Feb 26 #2/11 95.2%

HMMT February 2026 is a math competition benchmark based on problems from the Harvard-MIT Mathematics Tournament, testing advanced mathematical problem-solving and reasoning.

MathArena Apex #1/8 90.2%

MathArena Apex is a challenging math contest benchmark featuring the most difficult mathematical problems designed to test advanced reasoning and problem-solving abilities of AI models. It focuses on olympiad-level mathematics and complex multi-step mathematical reasoning.

IMO-AnswerBench #5/17 89.8%

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

MMLU-Pro #6/39 87.5%

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.

LiveBench #15/27 73.6%

LiveBench is a challenging, contamination-limited LLM benchmark that addresses test set contamination by releasing new questions monthly based on recently-released datasets, arXiv papers, news articles, and IMDb movie synopses.

Humanity's Last Exam #24/69 48.2%

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.

CodeForces #3/8 1

A competitive programming benchmark using problems from the CodeForces platform. The benchmark evaluates code generation capabilities of LLMs on algorithmic problems with difficulty ratings ranging from 800 to 2400.

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

⚡ DeepSeek-V4-Pro (1.6T MoE) Interactive VRAM & Token Speed Calculator

Concurrency (Batch):
MoE 1600B / A49B·61L (d=7168)·128Q/1KV·Ctx 1M
Runnable Models →
GPUs
96 GB VRAM·4000 GB/s·Safe Usable 94.1 GB
Out of VRAM (OOM)990.02 / 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)
990.02/ 94.1 GB usable

Physical 96 GB · Load 1052.3%

Theoretical Decode Speed
—tokens/s

Exceeds usable VRAM by 895.94 GB

Max Context at Batch=1
0K/ 1M

MLA compresses multi-batch KV footprint

Weights: 970 GBKV Cache (1×16K): 0.17 GBOverhead: 19.85 GB
⛔ Cannot Load
Reference Launch CLI:
OLLAMA_CONTEXT_LENGTH=16384 ollama run deepseek-v4:pro

Insufficient Local VRAM? Ways to Run DeepSeek-V4-Pro (1.6T MoE)

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

📊 DeepSeek-V4-Pro (1.6T MoE) Quantization VRAM Breakdown (16K Context)

Quantization Bits/Weight Weights KV Cache (16K) Total VRAM Quality Note
FP16 / BF16 16 bpw 3200 GB 0.17 GB 3264.62 GB Lossless Full Precision
FP8 / Q8_0 8.5 bpw 1700 GB 0.17 GB 1734.62 GB 99.9% Near-Lossless
GGUF Q6_K 6.56 bpw 1312 GB 0.17 GB 1338.86 GB Ultra High Quality
GGUF Q5_K_M 5.69 bpw 1138 GB 0.17 GB 1161.38 GB High Balanced Quality
Q4_K_M / AWQ 4.85 bpw 970 GB 0.17 GB 990.02 GB Sweet Spot (Recommended)
GGUF IQ4_XS 4.25 bpw 850 GB 0.17 GB 867.62 GB Compact 4-bit
GGUF Q3_K_M 3.91 bpw 782 GB 0.17 GB 798.26 GB Moderate Perplexity Loss
GGUF Q2_K 3.15 bpw 630 GB 0.17 GB 643.22 GB Extreme Compression

🖥️ GPU & Apple Silicon Compatibility for DeepSeek-V4-Pro (1.6T MoE) (Q4_K_M)

Frequently Asked Questions

What are the exact architecture parameters of DeepSeek-V4-Pro (1.6T MoE)?

DeepSeek-V4-Pro (1.6T MoE) uses DeepseekV4ForCausalLM with 61 layers (61-Layer CSA + HCA Hybrid Sparse Attention MoE + mHC (384 Routed + 1 Shared Expert)), hidden_size=7168, 128 Q heads / 1 KV heads (head_dim=512), and vocab_size=129,280.

How much VRAM does DeepSeek-V4-Pro (1.6T MoE) require?

At 4-bit quantization (Q4_K_M, 16K context), DeepSeek-V4-Pro (1.6T MoE) requires 990.02 GB VRAM. Full FP16 requires 3264.62 GB.