It's that time of year again. Google dropped Gemini 3. Anthropic released Claude Sonnet 4.5. OpenAI countered with GPT-5.2. And now, right on schedule, Chinese AI models are back in the conversation.

This happens every cycle.

After the American labs make their announcements, Chinese models quietly capture mindshare at year-end or early in the new year.

DeepSeek recently released their latest models, and they're flying somewhat under the radar so today we’re breaking down the DeepSeek ecosystem.

But first, let's catch up on what's happening in AI.

1. Circleback: AI-powered meeting notes and action item tracker

Auto-generates organized notes with speaker identification across 100+ languages, assigns follow-ups automatically, and searches past conversations using natural language queries. Integrates with Slack, Notion, HubSpot, Salesforce, and 10+ other apps to automate post-meeting workflows while maintaining SOC 2 Type II, HIPAA, and GDPR compliance.

2. Mocha: Build full-stack web apps using natural language prompts

Fully integrated platform with built-in auth, database, backend, and hosting. Google OAuth works out of the box, dev/production databases separate automatically, and you own all exported code with deployment taking just one click.

3. OpenCode: Open source AI coding agent now on desktop

Works with 75+ LLM providers including free models, Claude Pro, GPT, and Gemini with automatic LSP loading and multi-session support. Stores zero code or context data for privacy-sensitive environments.

DeepSeek is a Chinese AI company backed by High-Flyer, a quantitative hedge fund.

It operates more like a research side project than a traditional AI lab - they're focused on pushing capabilities forward rather than building polished consumer products.

The company is based in Hangzhou and has been releasing models since 2023.

What makes them notable: open weights, transparent research, and surprisingly strong performance at significantly lower costs than Western competitors.

They don't have the ecosystem polish of OpenAI or Google.

Like ChatGPT's 500 model variants and confusing naming schemes, DeepSeek's lineup is a mess. You'll see V3, V3.1, V3.2, V3.2-Exp, V3.2-Thinking, V3.2-Speciale, R1, R1-0528, DeepSeek Coder, DeepSeek Math... it goes on.

Let’s narrow it down to what’s important.

Despite the long list, DeepSeek has two modern flagship lines:

DeepSeek V3.2 (general purpose)

Your everyday AI assistant.

Use it for writing, general coding, knowledge tasks, standard queries. Think ChatGPT 4o or Claude Sonnet.

All those variants (V3.2-Exp, V3.2-Thinking, V3.2-Speciale) are just tuning adjustments.

V3.2-Exp is experimental, V3.2-Thinking emphasizes chain-of-thought, V3.2-Speciale is the polished version.

Just use V3.2.

DeepSeek R1 (Reasoning-focused)

For complex math, logic problems, hard coding challenges, step-by-step reasoning. Think ChatGPT 5-Thinking or Claude Opus 4.5.

R1-0528 is the latest revision. It's slower but shows transparent reasoning - you see how it thinks through problems.

We ran three practical tests to see how these models perform.

Test 1: Coding - Reddit Scraper

Prompt: "Create a Python script that scrapes the top 10 posts from a subreddit and saves them to a CSV file. The CSV should include: post title, author, upvotes, and URL. Include error handling and comments explaining each step."

Model: DeepSeek R1

Result:

DeepSeek over-engineered this into an interactive CLI application with pandas when the prompt asked for a simple script.

It added features nobody requested - timestamps, comment counts, interactive menus.

And it went with the hacky JSON endpoint approach instead of using PRAW (Reddit's official API), which is the correct engineering decision.

Gemini 3 nailed it - PRAW library, built-in csv module, clean edge case handling for deleted users.

Takeaway: The right answer is often the simplest one. Gemini understood the assignment. DeepSeek didn’t.

Test 2: Problem-Solving - Equity Split

Prompt: "A startup has 3 co-founders who need to split equity. Founder A contributed the initial idea and worked full-time for 6 months. Founder B joined 3 months ago as CTO with critical technical expertise. Founder C is a part-time advisor who will become full-time after funding. Design an equitable split accounting for time commitment, role importance, future contributions, and vesting schedules. Show your reasoning step-by-step."

Model: DeepSeek R1

Result:

DeepSeek used a "Past vs Future" weighting (50/50) that fundamentally doesn't make sense for startups.

Deepseek’s output reads like stream-of-consciousness math - "Let's allocate... Actually, wait, better to think..." - which kills confidence in the recommendation.

Gemini 3 provided strategic advice with specific legal mechanisms: advisor shares vs co-founder shares, double-trigger acceleration, conditional grants.

Takeaway: DeepSeek treated this like a homework assignment. Gemini treated it like startup advice.

DeepSeek's reasoning transparency doesn't help if the reasoning itself is flawed.

Test 3: Writing - LinkedIn Post

Prompt: "Write a LinkedIn post (max 200 words) announcing 'AI Dev Weekly' newsletter covering: Latest AI model releases and benchmarks, practical implementation guides, code examples and API comparisons, open-source tool reviews. Tone: Professional but approachable, technical but not jargon-heavy. Goal: Get developers to subscribe."

Model: DeepSeek V3.2

Result:

DeepSeek wrote a functional product description.

It's concise, hits the keywords, but reads corporate and promotional.

Gemini 3 used Problem-Agitation-Solution framework, opened with a contrasting belief ("Building > Reading"), and wrote peer-to-peer instead of brand-to-consumer.

It understands social psychology.

Takeaway: This is where you see the ranking difference. Gemini 3 is top-tier for a reason - it gets human psychology and platform dynamics.

DeepSeek produced what an AI should produce.

Honest verdict:

Across all three tests: DeepSeek isn't bad. But it's nowhere near the top models.

Gemini 3 won all three tests.

The LMSys rankings exist for a reason. DeepSeek at #8-10 and the results here validate those rankings well.

Rankings: On paper, DeepSeek isn't the strongest.

DeepSeek V3 ranks 8th on LMSys Arena (Elo: 1373), R1 ranks 10th (Elo: 1358). Gemini 3 and GPT-5.2 dominate the top spots.

But:

Open weights - You can download, modify, fine-tune for custom applications

Cost efficiency - API pricing is significantly lower ($0.27-0.55 per million tokens vs $2.50-7.50 for competitors)

Strong technical performance - Punches above its ranking for coding and math

The caveat: It's a Chinese model..

DeepSeek is just an alternative tool that’s open, transparent and competent. In terms of output it’s not standing out.

Reply below: What would you use DeepSeek for? And do you trust it?

Until next time,
Vaibhav 🤝🏻

If you read till here, you might find this interesting

#AD 1

Your dream retirement isn’t going to fund itself—that’s what your portfolio is for.

When generating income for a comfortable retirement, there are countless options to weigh. Muni bonds, dividends, REITs, Master Limited Partnerships—each comes with risk and oppor-tunity.

The Definitive Guide to Retirement Income from Fisher investments shows you ways you can position your portfolio to help you maintain or improve your lifestyle in retirement.

It also highlights common mistakes, such as tax mistakes, that can make a substantial differ-ence as you plan your well-deserved future.

Download your free copy today.

#AD 2

Sure, if you held your stocks since the dotcom bubble, you would’ve been up—eventually. But three years after the dot-com bust the S&P 500 was still far down from its peak. So, how else can you invest when almost every market is tied to stocks?

Lo and behold, billionaires have an alternative way to diversify: allocate to a physical asset class that outpaced the S&P by 15% from 1995 to 2025, with almost no correlation to equities. It’s part of a massive global market, long leveraged by the ultra-wealthy (Bezos, Gates, Rockefellers etc).

Contemporary and post-war art.

Masterworks lets you invest in multimillion-dollar artworks featuring legends like Banksy, Basquiat, and Picasso—without needing millions. Over 70,000 members have together invested more than $1.2 billion across over 500 artworks. So far, 25 sales have delivered net annualized returns like 14.6%, 17.6%, and 17.8%.*

Want access?

[Click here to skip the waitlist →]

_{Investing involves risk. Past performance not indicative of future returns. Reg A disclosures at masterworks.com/cd}