Bonus: Top up now and we'll double your first deposit — get x2 credits instantly.
Get Started
Sign Up
Sign Up

DeepSeek V3.2: Self-Hosting & Deployment Guide

DeepSeek V3.2 is a large open-source Mixture-of-Experts language model from DeepSeek that emphasizes high reasoning performance and efficient long‑context inference.

Browse available models

What is DeepSeek V3.2?

DeepSeek V3.2 is a cutting-edge open-source Mixture-of-Experts large language model by DeepSeek, with around 685B total parameters and ~37B active parameters per token that targets GPT-5-class reasoning and agent performance. It is primarily used for advanced reasoning and agentic tool-use workflows, such as long-horizon automation, complex planning, and multi-step decision-making in production environments. It is also widely used for long-context coding assistance, code generation and debugging, as well as large-document and RAG-style analysis thanks to context windows on the order of 128K–160K tokens. As its name suggests, DeepSeek V3.2 belongs to the DeepSeek V3 family and succeeds earlier DeepSeek V3.x experimental variants as a frontier open-weight model.

How to run DeepSeek V3.2 yourself

Open weights (mit), 685.4B parameters. Below: how much memory the weights need at each quantization, which hardware fits them, and the commands to serve it.

DeepSeek V3.2 VRAM requirements

Precision / quantizationWeightsFigureSource
BF16 (full precision)1370.7 GBEstimate685.4B params × 2 bytes
AWQ 4-bit399.8 GBFile sizeQuantTrio/DeepSeek-V3.2-Exp-AWQ
MLX 4-bit378.1 GBFile sizemlx-community/DeepSeek-V3.2-4bit
GGUF Q8_0713.3 GBFile sizeunsloth/DeepSeek-V3.2-GGUF
GGUF Q4_K_M405.4 GBFile sizeunsloth/DeepSeek-V3.2-GGUF
GGUF Q2_K245.4 GBFile sizeunsloth/DeepSeek-V3.2-GGUF

File size = actual download size on Hugging Face. Estimate = parameter count × bytes per weight. Weights only — the KV cache for long contexts needs extra memory; the fit tables below add 15% headroom as a rule of thumb.

GPU requirements: deploy DeepSeek V3.2 on your hardware

HardwareBF16 (full precision)AWQ 4-bitMLX 4-bitGGUF Q8_0GGUF Q4_K_MGGUF Q2_K
NVIDIA H200141 GB—4×4×6×4×3×
NVIDIA H10080 GB—6×6×—6×4×
NVIDIA A100 80GB80 GB—6×6×—6×4×
NVIDIA L40S48 GB—————6×

GPUs needed per variant.

HardwareGGUF Q8_0GGUF Q4_K_MGGUF Q2_K
RTX 509032 GB———
RTX 409024 GB———
RTX 309024 GB———

GGUF via llama.cpp; more cards or CPU offload for larger files.

HardwareGGUF Q8_0GGUF Q4_K_MGGUF Q2_KMLX 4-bit
Mac, 512 GBM3 Ultra——Fits—
Mac, 192 GBM2 Ultra————
Mac, 128 GBM4 Max / M3 Max————
Mac, 64 GBM4 Pro / Max————

Assumes ~75% of unified memory is usable by the GPU.

HardwareBF16 (full precision)AWQ 4-bitMLX 4-bitGGUF Q8_0GGUF Q4_K_MGGUF Q2_K
AWS p5.48xlarge8× H100 · 640 GB—FitsFits—FitsFits
AWS p4de.24xlarge8× A100 80GB · 640 GB—FitsFits—FitsFits
AWS g6e.12xlarge4× L40S · 192 GB——————
Google Cloud a3-highgpu-8g8× H100 · 640 GB—FitsFits—FitsFits
Azure ND H100 v58× H100 · 640 GB—FitsFits—FitsFits

Whole-instance GPU memory; public instance specs.

Serve it yourself

pip install -U vllm
vllm serve deepseek-ai/DeepSeek-V3.2
pip install 'sglang[all]'
python -m sglang.launch_server --model-path deepseek-ai/DeepSeek-V3.2
llama-server -hf unsloth/DeepSeek-V3.2-GGUF:Q4_K_M \
  --ctx-size 32768 --n-gpu-layers 99
pip install -U mlx-lm
mlx_lm.generate --model mlx-community/DeepSeek-V3.2-4bit --prompt "Hello"

Commands use the official repositories above; vLLM and SGLang expose an OpenAI-compatible endpoint.

Not on LLM.API yet

We do not route this model through our API at the moment. Browse the models you can call today — most workloads have a close match already live.

Browse available models

Providers

Different companies host the same model. OpenRouter routes your request to one of them based on the routing mode you pick — Balanced (price + speed), Nitro (fastest), or Exacto (highest tool-calling accuracy).

List price by provider

InputOutput
zyphra$0.25 in
$0.5 out

Provider list prices; the LLM.API discount applies on top.

Provider Input Output Cache read /M Latency Throughput Uptime
zyphra30% offin $0.25; out $0.5 per 1M tokens164K tokenstools, streaming, JSON, structured

Prices and availability from the LLMAPI catalogue, updated nightly. Last updated 27 Sept 2026.

Try this model

Test DeepSeek V3.2 right here — free to start.

DeepSeek V3.2
Hi! Want to test the model?

Suggestions for your first prompt

Not available on LLM API yet

We do not route this model through our API at the moment, so there is no endpoint or code snippet for it yet. Browse the models you can call today — most workloads have a close match already live.

Browse available models

DeepSeek V3.2: Self-Hosting & Deployment Guide in production

What the model is good at.

  • Conversational Chat

    Engages in multi-turn dialogue, following instructions, answering questions, and maintaining context for general-purpose conversational assistance.

  • Document Reasoning

    Analyzes and summarizes long-form text, extracting key points, performing reasoning, and answering questions based on provided content.

  • Multilingual Translation

    Translates between multiple languages while attempting to preserve meaning, style, and domain-specific terminology across diverse text inputs.

  • Visual Understanding

    Interprets images to identify objects, scenes, relationships, and described content for downstream reasoning or question answering tasks.

  • Text OCR Extraction

    Reads and extracts textual content from images, including documents, screenshots, or signs, enabling downstream search, analysis, and transformation.

What you get when you call DeepSeek V3.2: Self-Hosting & Deployment Guide through LLM.API.

  • Unified AI Routing

    Dynamically route each request to the optimal model across providers based on latency, cost, and quality—without changing your app code or client integration.

    One endpoint, any model

  • Cost-Aware Orchestration

    Automatically balance cost and performance with configurable policies that pick cheaper models for routine calls and premium models only when they truly matter.

    Spend less per token

  • Intelligent Fallbacks

    Configure per-route failover to backup models and providers so outages, rate limits, or timeouts don’t take your AI features offline.

    Resilient by default

  • Deep Observability

    Get per-request traces, latency, cost, and model metrics across all providers in one place, with logs ready for debugging and optimization.

    See every token

  • Task-Level Abstractions

    Define high-level tasks like chat, retrieval, or tools once and let LLM.API handle model-specific prompts, parameters, and orchestration behind a stable contract.

    Code to tasks, not models

  • High-Throughput Batch

    Submit massive batches through a single API to parallelize inference, slash per-request overhead, and unlock bulk processing workflows at scale.

    Throughput at scale

When to Use — When NOT to Use

Use it if...

  • You need a cost-effective general-purpose model for everyday coding and content tasks.
  • You need decent multilingual understanding and translation without requiring best-in-class quality.
  • Your use case involves batch-processing many small requests where price sensitivity is critical.
  • You need a capable assistant for code explanation, minor refactors, and simple bug hunting.
  • Your use case involves lightweight data extraction or summarization from short to medium texts.
  • You need a backup or secondary model to diversify providers for resilience and cost.

Avoid if...

  • You need frontier-level reasoning performance on complex, multi-step scientific or mathematical problems.
  • Your workload requires highly reliable compliance, safety filters, and mature enterprise governance tooling.
  • You need deeply specialized domain knowledge validated against cutting-edge research or proprietary standards.
  • Your workload requires tightly integrated ecosystem tools, plugins, or advanced function-calling capabilities.
  • You need proven, battle-tested performance at very large context windows for lengthy documents.
  • Your workload requires strict SLAs, global support guarantees, and long-term enterprise stability assurances.

LLM API rating for DeepSeek V3.2

★ 4.0/ 5.0LLM API rating
50% production adoption · 30% category benchmark · 20% tooling & value✓ Verified benchmark · LMArena + Artificial Analysis
4.8Production adoptionHighest-traffic open-weight family
3.0Benchmark qualityArtificial Analysis Intelligence Index 21.5 · LMArena text rank #99, Elo 1424
3.6Developer ecosystemContext window, tool and JSON support, price against the category median
Reasoning & chat 4.0Coding 4.1Vision & multimodal 4.1Tool & function calling 4.5
How is this rating calculated?

Every model gets a separate 1.0–5.0 score for each category it can be used in, so a model can rate high for coding and lower for vision. The score combines three parts:

  1. Production adoption — 50%. How widely the model is actually used in production: provider traffic, LiteLLM and gateway usage, open-weight downloads and its LMArena rank. Proven, widely used models score highest.
  2. Category benchmark — 30%. The public benchmark that matters for that category — the Artificial Analysis Intelligence Index and LMArena Elo for reasoning, LMArena coding Elo for coding, image and document support plus vision Elo for vision, tool and JSON support for function calling, measured speed for voice.
  3. Tooling & value — 20%. Context window, support for tools, structured outputs and streaming, and price compared with the category median.

Scores of 4.0 and above are shown in green, 3.0–3.9 in yellow, below 3.0 in red. When a model has no public data for a part, it gets a neutral value instead of an invented number.

A “Verified benchmark” mark (✓) means an independent lab — LMArena or Artificial Analysis — has actually tested that model, and the source figure is shown. Models without that mark are labelled “Estimated from published specs”: the score comes from the published context window, tool and schema support, price and adoption only, because no independent lab has tested them yet.

One exception: a brand-new top-tier model that scores very high on the independent benchmarks but has no usage history yet is rated on the benchmark (60%) and tooling & value (40%) alone, so a missing adoption figure can't hold it back. All inputs come from public sources and are refreshed weekly.

Adoption and benchmark inputs come from public sources — LMArena, Artificial Analysis Intelligence Index and published model specs. Updated weekly.

COMMUNITY

What developers say about DeepSeek V3.2

LMArena#99 · Elo 1424Arena coding#96 · Elo 1449Speed—Sources: LMArena, Artificial Analysis

Summarised from public developer discussions and write-ups. Opinions are the sources’, not LLM.API’s.

  • Developers on Greater Wrong report that DeepSeek V3.2 is a strong open-source model for mathematics and pricing, though it underperforms benchmarks in practice and feels slow compared to frontier competitors.
  • Developers on GitHub and Hugging Face have noted significant chat template changes in V3.2, specifically regarding new formats for tool calling and the introduction of 'thinking with tools' capabilities.
  • Reviews on ChatForest highlight that the introduction of DeepSeek Sparse Attention (DSA) in V3.2 effectively reduces computational complexity and API costs for long-context processing by 50% compared to V3.1.

SOURCES

Frequently Asked Questions

  • What is DeepSeek V3.2?

    DeepSeek V3.2 is a general-purpose large language model by DeepSeek focused on code, reasoning, and tool-using capabilities.

  • What is DeepSeek V3.2 best suited for?

    DeepSeek V3.2 is best for code generation, step-by-step reasoning, data transformation, and building chat-style assistants with strong instruction-following.

  • What is the context window of DeepSeek V3.2?

    DeepSeek V3.2 supports a context window up to 32K tokens, suitable for long conversations and larger documents.

  • How fast is DeepSeek V3.2 when called through LLM.API?

    Typical end-to-end latency ranges from a few hundred milliseconds to a few seconds, depending on prompt size and requested output length.

  • What modalities does DeepSeek V3.2 support via LLM.API?

    Through LLM.API, DeepSeek V3.2 currently supports text input and text output only.

  • How is DeepSeek V3.2 priced on LLM.API?

    LLM.API bills DeepSeek V3.2 per input and output token, with exact rates specified in the LLM.API pricing documentation.

  • How do I call DeepSeek V3.2 from the LLM.API endpoint?

    Specify the model name "deepseek-v3.2" (or the exact identifier from LLM.API docs) in your API request's model parameter.

  • How does DeepSeek V3.2 compare to similar models on LLM.API?

    DeepSeek V3.2 generally targets a balance of reasoning quality and cost, often being cheaper than top-tier frontier models with comparable capabilities.

  • Does DeepSeek V3.2 support tools or function calling via LLM.API?

    Yes, if enabled by LLM.API, DeepSeek V3.2 can consume tool definitions and output structured tool call arguments.

  • What are the main limitations of DeepSeek V3.2?

    DeepSeek V3.2 can hallucinate facts, lacks real-time knowledge, and may struggle with highly domain-specific or very long multi-step tasks.

COMPARE

Competitive Models

DeepSeek V4 Flash is an efficiency-optimized Mixture-of-Experts large language model from DeepSeek, featuring a 1M-token context window and fast inference for high-throughput applications.

  • DeepSeek V4 Pro

    Up to 30%
    • Instruction Following

    DeepSeek V4 Pro is DeepSeek’s flagship open-weights Mixture-of-Experts language model with a 1 million token context window and strong reasoning and coding capabilities. It is notable for combining frontier-level performance with open licensing and relatively low-cost deployment options.

  • DeepSeek V4 Flash (free)

    Up to 30%
    • Instruction Following

    DeepSeek V4 Flash (free) is an open-source, efficiency-optimized Mixture-of-Experts language model from DeepSeek, offering a 1M-token context window with only 13B parameters activated per token out of 284B total. It is designed to deliver fast, cost-effective long-context reasoning, coding, and agentic workflows.

    • Get one key to every model

    Swap your API key. Keep your code.

    Browse available models Talk to sales