DeepSeek-V4-Flash (284B-A13B MoE) Architecture Parameters, Benchmarks & VRAM
DeepSeek-V4 efficient MoE (284B total / 13B active) with 1M context window and compressed sparse/MLA attention.
🧬 DeepSeek-V4-Flash (284B-A13B MoE) Architecture & Tensor Specifications
Verified directly against HuggingFace (deepseek-ai/DeepSeek-V4-Flash) config.json & Model Card
1. Backbone & Dimensions
deepseek_v4- Total / Active Params
- 284B / 13B
- Hidden Size (d_model)
- 4,096
- MoE Expert Dim
- 2,048
- MoE Experts
- 256 Routed · Top-6 + 1 Shared
- Activation / Norm
- SiLU (SwiGLU) · ε=1e-6
- Native Tensor Dtype
- FP4 + FP8 Mixed
2. Attention & Layer Layout
MLA- Num Hidden Layers
- 43 Layers
- Hidden Layer Layout
- 43-Layer Hybrid CSA + HCA Compressed Sparse Attention MoE + mHC
- Attention Heads (Q / KV)
- 64 Q / 1 KV (64:1 GQA)
- Head Dimension
- 512
- Linear / KV Compression
- CSA + HCA Hybrid Attention: uses only 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 1M Context)
- Vocab Size
- 129,280 (Untied)
- Special Token IDs
- BOS: 0 · EOS: 1
📋 DeepSeek-V4-Flash (284B-A13B MoE) Official API & System Card Specifications
Verified against official DeepSeek developer documentation & pricing
1. Identity & Capabilities
API Spec- API Model ID
- deepseek-v4-flash
- Release Date
- 2026-07-31
- 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
CNY / 1M- Input / 1M Tokens
- ¥ 3
- Cached Input / 1M
- ¥ 0.1 -97%
- Output / 1M Tokens
- ¥ 9
- Tiered Pricing Notes
- Volcengine Ark: Volcengine Ark high-concurrency CN endpoint
🏆 Official Benchmark Results
Sourced from official System Card / Model Card & technical reports
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.
Chinese short-form factual accuracy benchmark
PhD-level Google-proof science reasoning questions
Humanity's Last Exam frontier academic benchmark
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.
LiveCodeBench is a holistic and contamination-free evaluation benchmark for large language models for code.
Scaled Model Context Protocol multi-tool coordination
🎛️ Official Sampling & Reasoning Parameters
Recommended generation settings & API parameter constraints
💡 Set temperature=1.0, top_p=1.0; allocate at least 384K context window for Think Max mode.
📊 DeepSeek-V4-Flash (284B-A13B MoE) Full Benchmark Scorecard
19 public benchmarks · ranks among 154 models · synced 2026-11-20
💻 Coding 6 items
LiveCodeBench is a holistic and contamination-free evaluation benchmark for large language models for code.
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.
A multilingual benchmark for issue resolving in software engineering that covers Java, TypeScript, JavaScript, Go, Rust, C, and C++.
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 is an updated benchmark for testing AI agents' tool use ability to operate a computer via terminal.
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.
🤖 Agents & Tools 2 items
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.
Tool Decathlon is a comprehensive benchmark for evaluating AI agents' ability to use multiple tools across diverse task categories. It measures proficiency in tool selection, sequencing, and execution across ten different tool-use scenarios.
📐 Math 6 items
HMMT February 2026 is a math competition benchmark based on problems from the Harvard-MIT Mathematics Tournament, testing advanced mathematical problem-solving and reasoning.
IMO-AnswerBench is a benchmark for evaluating mathematical reasoning capabilities on International Mathematical Olympiad (IMO) problems, focusing on answer generation and verification.
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.
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.
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.
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.
📜 Long Context 2 items
MRCR 1M is a variant of the Multi-Round Coreference Resolution benchmark designed for testing extremely long context capabilities with approximately 1 million tokens. It evaluates models' ability to maintain reasoning and attention across ultra-long conversations.
CorpusQA 1M is a long-context question answering benchmark designed to evaluate models at approximately 1 million token contexts. Models are scored on accuracy when retrieving and reasoning over information distributed across an extremely long input corpus.
🧠 Reasoning & Science 2 items
A challenging dataset of 448 multiple-choice questions written by domain experts in biology, physics, and chemistry. Questions are Google-proof and extremely difficult, with PhD experts reaching 65% accuracy.
SimpleQA is a factuality benchmark developed by OpenAI that measures the short-form factual accuracy of large language models. The benchmark contains 4,326 short, fact-seeking questions that are adversarially collected and designed to have single, indisputable answers.
Source: ZeroEval / llm-stats.com (✓ = third-party verified, others vendor-reported).
⚡ DeepSeek-V4-Flash (284B-A13B MoE) Interactive VRAM & Token Speed Calculator
Physical 512 GB · Load 38.2%
MoE: reads only 7.88 GB active weights/step
MLA compresses multi-batch KV footprint
OLLAMA_CONTEXT_LENGTH=16384 ollama run deepseek-v4:flashMac Studio M5 Ultra 512G Fits DeepSeek-V4-Flash (284B-A13B MoE) · Cost & TCO
Approx. $9,499 (~$19/GB VRAM) · 190W TDP
📊 DeepSeek-V4-Flash (284B-A13B MoE) Quantization VRAM Breakdown (16K Context)
| Quantization | Bits/Weight | Weights | KV Cache (16K) | Total VRAM | Quality Note |
|---|---|---|---|---|---|
| FP16 / BF16 | 16 bpw | 568 GB | 0.12 GB | 579.93 GB | Lossless Full Precision |
| FP8 / Q8_0 | 8.5 bpw | 301.75 GB | 0.12 GB | 308.35 GB | 99.9% Near-Lossless |
| GGUF Q6_K | 6.56 bpw | 232.88 GB | 0.12 GB | 238.11 GB | Ultra High Quality |
| GGUF Q5_K_M | 5.69 bpw | 202 GB | 0.12 GB | 206.6 GB | High Balanced Quality |
| Q4_K_M / AWQ | 4.85 bpw | 172.18 GB | 0.12 GB | 176.19 GB | Sweet Spot (Recommended) |
| GGUF IQ4_XS | 4.25 bpw | 150.88 GB | 0.12 GB | 154.46 GB | Compact 4-bit |
| GGUF Q3_K_M | 3.91 bpw | 138.81 GB | 0.12 GB | 142.15 GB | Moderate Perplexity Loss |
| GGUF Q2_K | 3.15 bpw | 111.83 GB | 0.12 GB | 114.63 GB | Extreme Compression |
🖥️ GPU & Apple Silicon Compatibility for DeepSeek-V4-Flash (284B-A13B MoE) (Q4_K_M)
Frequently Asked Questions
What are the exact architecture parameters of DeepSeek-V4-Flash (284B-A13B MoE)?
DeepSeek-V4-Flash (284B-A13B MoE) uses DeepseekV4ForCausalLM with 43 layers (43-Layer Hybrid CSA + HCA Compressed Sparse Attention MoE + mHC), hidden_size=4096, 64 Q heads / 1 KV heads (head_dim=512), and vocab_size=129,280.
How much VRAM does DeepSeek-V4-Flash (284B-A13B MoE) require?
At 4-bit quantization (Q4_K_M, 16K context), DeepSeek-V4-Flash (284B-A13B MoE) requires 176.19 GB VRAM. Full FP16 requires 579.93 GB.