If you run a brokerage, manage a rental portfolio, or build PropTech, you have probably already tried ChatGPT for listing copy and given up the moment you needed to feed it a 60-page lease, a stack of comps, or a multilingual buyer’s email thread. That is where DeepSeek for real estate becomes interesting. DeepSeek charges $0.14/million tokens input and $0.28/million tokens output for Flash, and $1.74/million input and $3.48/million output for Pro, with a one-million-token context window on both tiers — enough to drop in an entire lease bundle and ask precise questions. This guide walks through ten concrete workflows, the prompts that actually work, the limits you should know about, and what an honest monthly bill looks like at portfolio scale.

Why agents and PropTech teams are testing DeepSeek

Real estate is a document-heavy, data-heavy, deadline-driven business. The work that LLMs are good at — summarising long text, drafting in a consistent voice, pulling structured fields from messy inputs — maps almost one-to-one onto the daily grind of a listing agent, a lease administrator, or an acquisitions analyst.

The market backdrop is real. According to a January 2026 survey by Delta Media covered by Inman, 97 percent of brokerage leaders now report that their agents are actively using AI, and McKinsey estimates AI could generate $110 to $180 billion in annual value for the U.S. real estate sector — but as of 2026, only 5% of commercial real estate firms that have started AI initiatives have achieved all of their program goals. The gap is execution, not aspiration.

DeepSeek’s appeal in this category is narrow but specific:

• Long context. A million-token window holds a full sale-leaseback file, an entire rent roll, or a year of tenant emails without chunking gymnastics.
• Cost. At V4-Flash rates, summarising a 50-page lease costs cents, not dollars.
• OpenAI compatibility. Existing scripts that call ChatGPT need a one-line change to point at DeepSeek.
• Open weights. If your compliance team won’t allow tenant data to leave your VPC, you can self-host the same model. Read more on the DeepSeek V4 page.

What you’re working with: V4-Pro and V4-Flash

Before the workflows, the model lineup. DeepSeek published the DeepSeek-V4 Preview on April 24, 2026 — a new open-weight Mixture-of-Experts series. Two models ship today. V4-Pro packs 1.6 trillion total parameters with 49 billion activated per token. V4-Flash is the efficient sibling at 284 billion total / 13 billion active. Both support native 1M context, and both are open weights.

For most real estate workflows, DeepSeek V4-Flash is the right default. Reach for DeepSeek V4-Pro only when you have a hard problem — a contested lease abstraction, a complex underwriting memo — where the benchmark lift justifies roughly 12× the output cost.

API basics every team should know

Whether you build a property chatbot or a back-office automation, chat requests hit POST /chat/completions, the OpenAI-compatible endpoint at https://api.deepseek.com. The API is stateless — your client must resend the conversation history with every request, unlike the web/app, which maintains session history for the user. DeepSeek also exposes an Anthropic-compatible surface against the same base URL.

Thinking mode is a request parameter, not a separate model ID. Set reasoning_effort="high" with extra_body={"thinking": {"type": "enabled"}} for tasks like multi-step underwriting or contradictory-clause spotting. The response then returns reasoning_content alongside the final content.

If you maintain older integrations: deepseek-chat and deepseek-reasoner currently route to deepseek-v4-flash in non-thinking and thinking modes respectively, and they will be retired on 2026-07-24 at 15:59 UTC. Migration is a one-line model= swap; the base URL is unchanged.

from openai import OpenAI

client = OpenAI(
    base_url="https://api.deepseek.com",
    api_key="YOUR_KEY",
)

resp = client.chat.completions.create(
    model="deepseek-v4-flash",
    messages=[
        {"role": "system", "content": "You are a real estate copy editor."},
        {"role": "user", "content": "Rewrite this listing in 120 words: ..."},
    ],
    temperature=1.5,
    max_tokens=400,
)
print(resp.choices[0].message.content)

Temperature note: DeepSeek recommends 0.0 for code/math, 1.0 for data analysis, 1.3 for general chat and translation, and 1.5 for creative writing. Listing copy lives at 1.5; lease extraction lives at 0.0. For a deeper walkthrough see our DeepSeek API getting started guide.

Ten real-estate workflows that work today

1. Listing descriptions that don’t sound like a chatbot

Feed the model the MLS field dump and a short brand-voice sample. Ask for three variants — “neighbourhood-led”, “feature-led”, “lifestyle-led” — at 120 words each. With temperature=1.5, V4-Flash produces copy that needs light editing, not rewriting.

Prompt skeleton:

“You are a copywriter for [brokerage]. Voice: warm, specific, no clichés. Banned: ‘nestled’, ‘boasts’, ‘stunning’. Property facts: {fields}. Write three 120-word descriptions. Cite only facts in the input.”

2. Lease abstraction at portfolio scale

Drop a full lease into the 1M-token window, ask for a JSON abstract. The honest caveat: JSON mode is designed to return valid JSON, not guaranteed. Include the word “json” plus a small example schema in the prompt, set max_tokens high enough that the response cannot be truncated, and handle the rare empty-content case in your code.

Fields to ask for: tenant, premises, commencement, expiry, base rent schedule, escalations, options, CAM treatment, assignment/sublet, restoration. Cross-check against your lease management system; do not trust a single pass for legally binding fields.

3. Comparative market analyses (CMA) drafts

Pull comps from your MLS or data provider into a CSV. Send the CSV plus the subject property to V4-Pro with thinking mode on. Ask for three price scenarios with the reasoning shown. The model is not a licensed appraiser — its output is a draft for the agent to review, not a substitute for one.

Where this fits the broader pattern: The most widely adopted AI use cases in real estate include automated valuation models, predictive maintenance, smart building energy optimisation, document and lease analysis, conversational AI for tenant experience, and generative content for marketing.

4. Tenant and buyer email triage

Wire the inbox into a small classifier prompt: maintenance, payment, lease question, showing request, complaint, other. Route accordingly. With cached system prompts, this costs near-zero — see the cost example below.

5. Translation for international buyers

The CA/UK/AU markets have heavy non-English buyer flow. Use V4-Flash at temperature=1.3 for translation. For complex legal text, escalate to V4-Pro thinking mode and pair with a human reviewer. Compare with our DeepSeek for translation notes.

6. Inspection-report summaries for buyers

Inspection PDFs run 30–80 pages. Prompt: “Summarise as Major / Moderate / Minor / Cosmetic with page references and likely repair-cost bands. Do not estimate prices you cannot justify from the report itself.”

7. Disclosure and HOA-document Q&A

Drop the HOA bundle into context, let the buyer ask plain-English questions: “Are short-term rentals allowed? What is the special assessment history?” Always cite the page in the answer, and warn users it is a research aid, not legal advice.

8. Buyer chatbots on the brokerage site

A streaming response (stream=true) over POST /chat/completions gives a snappy chat UX. Use DeepSeek context caching on the system prompt and listing catalogue so each user turn pays the cache-hit rate, not the miss rate.

9. Marketing assets — social, email, blog

Standard DeepSeek for marketing patterns apply: long-form blog draft, sliced into LinkedIn, Instagram captions, and a newsletter. Ask the model to flag any factual claim it can’t ground in the input.

10. Internal search across years of deal files

Combine V4 with a vector store (see the DeepSeek RAG tutorial) so analysts can ask “show me every deal where we flagged Phase II environmental issues” against ten years of memos.

What an honest monthly bill looks like

Take a mid-size brokerage running 1,000,000 small chat/triage calls a month: a 2,000-token system prompt (cached across calls), a 200-token user message (uncached), 300-token response, on deepseek-v4-flash:

The same workload on deepseek-v4-pro costs roughly $1,682 — about 10× more — because Pro’s output rate is $3.48/M vs Flash’s $0.28/M. For chatbot and triage work, Flash is the obvious pick. Use the DeepSeek pricing calculator to model your own ratios. Rates verified against DeepSeek’s V4 Preview release notes as of April 2026.

Where DeepSeek falls short for real estate

• Property valuation precision. A general-purpose LLM is not an AVM. AI-powered automated valuation models now achieve median error rates of 2.8 per cent, down from 10–15 per cent five years ago, but those numbers come from purpose-built models trained on MLS, tax, and transaction data. Use HouseCanary, Zillow’s API, or a similar AVM for the number, and use DeepSeek to write the narrative around it.
• Live market data. The model does not browse the MLS. You must supply current comps, rents, and inventory.
• Lease specifics. General-purpose AI like ChatGPT doesn’t understand commercial lease structures, rent reviews, or outgoings apportionment as native concepts. For everyday communication and content, it’s excellent. For complex lease administration or portfolio management, purpose-built tools are necessary. The same caveat applies to DeepSeek — pair it with a lease management system, don’t replace one.
• Photos. The V4 family is text-first. For property image analysis (room types, condition scoring), specialist computer-vision tools like Restb.ai still own this category.
• Regulated outputs. Anything touching fair-housing language, ADA-compliant alt text, or licensed appraisal needs a human in the loop and an audit trail.

Privacy and jurisdiction — read this before piping in tenant data

DeepSeek’s hosted API processes requests on servers subject to Chinese law. For US, UK, EU, and AU brokerages handling personal data of buyers, sellers, and tenants, that has implications. Italy’s Garante ordered a block on the consumer app in January 2025; several US states have restricted DeepSeek use on government devices. None of that prevents you using the API for non-personal data, but it does shape your architecture choices.

Two practical paths:

1. Use the hosted API for non-PII. Listing copy, marketing drafts, anonymised comp analysis. Strip names, emails, and unit numbers before the request.
2. Self-host the open weights. V4-Flash at 13B active parameters is within reach of a well-equipped server. See install DeepSeek locally and DeepSeek privacy for the trade-offs.

Honest alternatives by sub-task

For broader options see DeepSeek alternatives and our real-world applications hub.

Getting started — a 30-day plan

1. Week 1. Pick one workflow (listing descriptions is the lowest-risk start). Get an API key via our key-setup walkthrough. Build a prompt template and review 50 outputs against your style guide.
2. Week 2. Add lease summarisation against five known leases. Score each abstract against the source. Tune the prompt; do not loosen the validation.
3. Week 3. Stand up the brokerage chatbot in staging with caching. Measure cost per conversation and time-to-first-token.
4. Week 4. Decide hosted vs self-hosted based on data sensitivity. Document the policy. Roll out to one office or one team — not the whole firm.

Last verified: 2026-04-25. DeepSeek AI Guide is an independent resource and is not affiliated with DeepSeek or its parent company. Model IDs, pricing and API behaviour change; check the official DeepSeek documentation and pricing page before committing to a production decision.