---
title: 'DeepSeek vs Gemini 3.0 — Full Guide & Tutorial to DeepSeek V3.2'
source: 'https://youtube.com/watch?v=dPCajxRXgLY'
video_id: 'dPCajxRXgLY'
date: 2026-06-17
duration_sec: 0
---

# DeepSeek vs Gemini 3.0 — Full Guide & Tutorial to DeepSeek V3.2

> Source: [DeepSeek vs Gemini 3.0 — Full Guide & Tutorial to DeepSeek V3.2](https://youtube.com/watch?v=dPCajxRXgLY)

## Summary

DeepSeek V3.2 is a free, open-source AI model that rivals top premium models in math and reasoning benchmarks. The video guides users on how to use its web interface, showcasing its capabilities in writing, data analysis, and coding. The creator discusses the two available versions and provides a practical tutorial for integrating the tool into daily workflows.

### Key Points

- **The Free AI Challenger** [0:00] — DeepSeek V3.2 is a free AI model that scored 96% on the American Invitational Mathematics Examination (AIME) and matches or beats GPT-5 and Gemini 3 Pro on reasoning benchmarks.
- **Two Flavors of DeepSeek** [0:51] — DeepSeek V3.2 has two versions: Standard V3.2, an everyday workhorse for general tasks, and V3.2 Special, a high-compute model focused on deep reasoning and competition coding.
- **Unique Thinking Integration** [2:14] — DeepSeek V3.2 integrates thinking directly into tool use, preserving reasoning traces across multiple tool calls for better multi-step problem solving.
- **Quick Setup Process** [5:14] — Setup takes about 60 seconds: go to chat.deepseek.com, sign up with Google, and access the chat interface with 'Deep Think' toggle for reasoning mode.
- **Real-World Use Case: Writing** [7:41] — With 'Deep Think' mode enabled, DeepSeek creates detailed outlines for articles, breaking down the structure and audience needs before generating output.
- **File Handling for Data Analysis** [10:20] — DeepSeek can upload CSV/PDF files and return structured summaries, such as sales data insights or key takeaways, without coding.
- **Instant Code Generation** [11:02] — No coding knowledge is needed: DeepSeek can generate and preview playable games or web apps (like a snake game or to-do list) right in the chat window.
- **Handling Multi-Step Business Problems** [14:07] — Using 'Deep Think', DeepSeek reasons through a complex business prompt (growth plan) and delivers a structured output with specific action items.
- **DeepSeek vs. Premium Models** [15:36] — DeepSeek wins on cost (free), transparency (visible thinking), math reasoning, and open weights. ChatGPT and Gemini win on polish, integrations, and stability.
- **Effective Prompting Tips** [18:45] — Eight tips are provided: structure prompts with markdown, use 'think and answer' format, be hyper-specific, use personas, ask for multiple approaches, request self-checks, use placeholders, and iterate.

### Conclusion

DeepSeek V3.2 is a capable, free alternative to ChatGPT and Gemini for most daily tasks, especially writing, analysis, and coding. Its open-weight nature and transparent reasoning make it a strong choice for users who do not rely on premium ecosystems.

## Transcript

You're paying $20 a month for Gemini or
Chad GPT. Meanwhile, there's a free AI
model that just hit 96% on the toughest
math benchmark in the world, matches
GPT5 in reasoning, and won gold medals
at international coding competitions.
And I'm willing to bet most of you have
never even tried it. It's called
Deepseek V3.2. It dropped December 1st,
completely free, completely open source,
and it's shockingly good. Most people
forgot Deepseek even exists. That's a
mistake because in the next 20 minutes,
I'm going to show you how to use this
thing for real work, writing, research,
coding, data analysis, all without
touching a single line of code. You'll
see which version to use when thinking
mode actually matters and whether you
can finally ditch your Chat GPT or
Gemini subscription. So, let's go. So,
here's the situation. Back in January,
Deepseek made headlines with their R1
model. cheap to train, open source,
performed like the big boys. Then
everyone moved on. Open AAI launched
GPT5. Google dropped Gemini 3 Pro. The
hype cycle kept spinning and Deepseek
just kept building. On December 1st,
2025, they quietly released version 3.2
and the benchmarks are wild. 96% on the
American Invitational Mathematics
Examination. That's M, one of the
hardest math test on the planet. For
context, GPT5 High scored 94.6. Deepseek
beat it. On the Harvard MIT math
tournament, Deepseek hit 99.2%.
Gemini 3 Pro scored 97.5.
Again, Deep Seek wins. But here's the
thing. This isn't just one model.
Deepseek V3.2 comes in two flavors.
Standard V3.2, which is your everyday
workhorse. Fast, reliable GPT5 level
performance. Think of it as your Chad
GBT replacement. Then there's V3.2
Speciial. And Special is the beast.
Maxed out reasoning, deep chain of
thought. This thing won gold medals at
the International Mathematical Olympiad,
the International Olympiad in
Informatics, second place at the ICPC
World Finals. We're talking elite
competition level AI. And here's what
makes this release different. Deep Seek
V3.2 2 is the first model to integrate
thinking directly into tool use. Most AI
models lose their train of thought every
time they call an external tool. They
have to restart reasoning from scratch.
Deepseek preserves the reasoning trace
across multiple tool calls. That means
smoother workflows, better agent
performance, and way more reliable
multi-step problem solving. Now, there
is one catch with special. It's API only
right now and that API endpoint expires
December 15th, 2025. Why? Because it's
token hungry. To solve one code forces
problem, Special uses 77,000 tokens.
Gemini uses 22,000. The inference costs
are insane. So, Deep Seek is treating
Special as a research preview. After
December 15th, they'll likely roll it
into a more efficient production version
or just keep the standard model as the
main offering. But the standard V3.2 too
that's available right now free on the
web on mobile open source under an MIT
license no credit card no limits and it
performs at GPT5 level this isn't a toy
this is production ready AI let me break
down the two models so you know which
one to use standard V3.2 2. This is the
daily driver. It scored 93.1% on Emmy
2025. Code forces rating of 2386.
That puts it in the top 1% of
competitive programmers globally. It's
fast, it handles tool use, and it's free
via the web interface at
chat.deseek.com.
Use this for writing, brainstorming,
quick analysis, code, and tasks, and
anything you'd normally throw at chat
GPT. Then there's V3.2 special. This is
the high compute reasoning monster. 96%
on Amy, 99.2 on the Harvard MIT math
tournament, gold medals at IMO, IOI,
ICPC, and the Chinese math Olympiad.
Special is designed for deep reasoning,
multi-step proofs, complex research,
heavy analysis, but it doesn't support
tool calling right now. It's pure
thinking mode, and it's only available
through the API until December 15th. Why
can't you install Special locally yet?
It's in a research beta phase. The model
is publicly available on hugging phase,
but running it requires serious
hardware. 671
billion parameters total, but only 37
billion active per token thanks to the
mixture of experts architecture. You'd
need multiple high-end GPUs and
distributed compute to run it smoothly.
The standard V3.2 too. It's easier to
self-host if you've got the hardware,
but for most people, the web interface
is the way to go. Rule of thumb,
standard for speed, special for depth.
But since Special is temporary and API
gated, I'm going to focus this tutorial
on the standard V3.2 model, which is
what you can actually use today without
jumping through hoops. All right, let's
get you set up. This takes about 60
seconds. Step one, open your browser and
go to chat.deseek.com.
You'll see a simple landing page. Click
sign up. I just used the Google signin
option. One click. No password to
remember. Done. Step two. Once you're
in, you'll see the main chat interface.
The prompt box is right in the center.
Below it, you've got two toggles. One
says deep think. That enables extended
reasoning mode where the AI thinks
through your problems step by step. The
other toggle is search, but as of right
now, it's not reliably working. When it
does work, it functions like Chad GBT's
web search, but don't rely on it yet.
You also see a file upload button.
Deepseek is multimodal. You can upload
text files, spreadsheets, PDFs, images.
It'll extract the text and work with it.
No video or audio support yet, but for
documents and data, it's solid. On the
left side, you've got your chat history.
You can rename or delete chats by
clicking the three dots next to each
one. There's also a QR code button to
download the mobile app, which I'll show
you in a second. Step three,
understanding the interface. In normal
mode, Deep Seek gives you a
straightforward answer. Simple, clean,
fast. But when you enable deep think,
something interesting happens. The AI
thinks for several seconds first. You
literally watch its thought process
unfold in real time. Then it delivers
the final answer. This is where you see
the reasoning chain, the trade-offs.
it's considering the logic behind its
response. Step four, enable deep think
when you need it. If you're doing
something simple like writing an email
or brainstorming ideas, leave it off.
The model responds faster. But if you're
working on complex problem solving,
research, or multi-step tasks, turn on
Deepthink. The model will take longer to
respond, but the quality goes way up.
Pro tip: Prompt in English or Chinese
only. Deepseek's training is optimized
for those two languages. Other languages
go through a translation layer and the
results can be hit or miss. If you need
the output in another language, prompt
in English first, then ask DeepSk to
translate the final answer. That's it.
You're set up. Now, let's put this thing
to work. This is where it gets
interesting. I'm going to show you five
real use cases. Some are practical, some
are just fun. All of them work without
coding, without APIs, and without local
installs. Just your browser and a
prompt. Let's start with writing. I'm
going to ask Deepseek to create a long
form article outline. I'll turn on deep
think so it thinks through the structure
prompt. Create a detailed outline for a
2,000word article on remote work
productivity. Include five main
sections, each with three sub points.
Target audience is freelancers and small
business owners. Make it actionable.
Watch what happens. Deepseek thinks
first, breaking down the prompt. It's
identifying the audience, thinking about
pain points, structuring the flow. Then
it delivers a clean, hierarchical
outline with specific subtopics. No
fluff, just structure. This is where
Deep Seek shines for content work.
You're not just getting an answer.
You're seeing the reasoning. And if you
don't like a section, you can ask it to
revise just that part without starting
over. Here's something cool. We've been
using Deep Seek for text, but what about
video? Because that's the other part of
content creation. most people struggle
with. I've been testing Luma AI's RA3
modify for the visuals in this video.
And the workflow is honestly
fascinating. It's not just type a prompt
and hope for the best. There is this
reasoning engine built in that actually
thinks through what you're trying to
create before it starts generating.
Here's how it works. You either type a
detailed prompt or upload an image as
your starting point. Ray 3 modify
analyzes your creative intent, figures
out the physics, lighting, motion paths,
all of that. Then it gives you two
generation modes. Draft mode, which is
incredibly fast and cheap. Use that to
experiment and iterate. Once you nail
the concept, you upscale to hi-fi 4K HDR
for your final render. But the real game
changer, it's visual annotation. Instead
of writing a 10-line prompt trying to
describe motion, just draw on the image.
Circle an object, draw an arrow showing
how you want it to move. Ray 3 modify
handles the rest. The physics, the
lighting transitions, the camera work,
it all just works. I tested this with a
product demo concept. Started in draft
mode, tweaked the motion a few times,
then rendered the final 4K version. The
output was clean enough to use straight
out of the box. no post-processing. And
compared to other AI video tools I've
tried, this one actually preserves
anatomy and handles complex motion
without the usual AI artifacts. If
you're building content, product videos,
or anything visual, Ray 3 modify is
worth trying. You can find it at
dreamachine.lumalabs.ai.
They give you credits to start so you
can test draft mode and visual
annotation yourself. Link below. All
right, back to Deep Seek. Now, let's
test file handling. I'm going to upload
a sample CSV file with sales data. Then
I'll ask Deepseek to analyze it. Prompt:
Analyze this sales data. Give me the top
three insights and identify any
concerning trends. Deepseek reads the
file, processes the data, and returns a
structured summary. It highlights that
Q3 revenue dropped 15%, identifies the
underperforming product category, and
flags a regional sales dip. All in plain
English. No pivot tables, no formulas,
just insights. You can also upload PDFs
and ask for summaries, key takeaways, or
specific extractions. For research work,
this is a massive timesaver. All right,
this is the showstopper. I'm going to
show you three things you can build in
Deep Seek right now. No coding knowledge
required. These are instant playable
visual demos. First, let's start with
something fun. generating a simple
JavaScript game and playing it right
inside DeepSseek prompt. Create a simple
snake game in HTML and JavaScript that I
can play right now. Deepseek writes the
code. A preview button appears. I click
it. Boom. The game loads in the
interface. I'm playing Snake inside the
AI chat window. No libraries, no
installs. 15 seconds from prompt to
playable game. Now, let's take it a step
further and build a simple space shooter
arcade game. Prompt: Build a simple
space shooter arcade game with HTML 5
canvas. Make it playable. Again, Deep
Seek generates the code. I hit preview.
The game loads. I'm firing shots at
targets. It works. Chad GPT would give
you the code and tell you to paste it
somewhere. Deep Seek lets you run it.
And now, let's switch gears and create a
simple to-do list app. Prompt. Create a
to-do list app with add and delete
functionality. Full web app, interactive
UI. I can add tasks, delete tasks, check
them off. It's live. It's functional.
And I didn't write a single line of
code. Now, before I move on to comparing
Deep Seek with CHG GBT and Gemini, let
me show you something that we've built.
While you're learning DeepS or any AI
tool for that matter, I can't stress
enough how helpful it is to have
everything in one place. That's exactly
why I built AI Master Pro. It's an
all-in-one AI hub where you're not just
watching tutorials. You're actually
applying what you learn in real time.
Here's what's inside. The AI Master
Method course which walks you through AI
foundations, workflows, and how to
actually sell AI services. Over a 100
lessons, templates, PDFs, everything you
need to go from beginner to building
real AI products in four weeks. But the
course is just the starting point. What
I actually use every single day are the
AI tools built right into the platform.
There's a personal AI master trained on
unique data that can teach you anything
about AI 24/7. You can generate
highquality prompts on the fly and with
integrated tools like Sora, VO, Nano
Banana directly into the platform. So if
you join before the end of 2025, you
will get bonus generation credits. You
also get Prompt Lab Pro, 300 plus
readyto-use prompts you can copy and
paste, discounts on other AI tools, and
a curated weekly AI digest so you never
fall behind on what's new, so you can
learn AI and use AI in the same place.
No tab switching, no trying to remember
which tool does what. Everything's
organized. Everything's accessible and
honestly, I'm on the platform every
single day. If you want to check it out,
we're offering a huge discount for the
first thousand people who join the
annual plan. Link below. All right, back
to Deepseek. Let's push it harder. I'm
going to give Deepseek a multi-step
business problem and watch how it
handles it. Prompt: I'm launching a
subscription software product. Price is
$49 per month. Target market is
freelancers. I need a six-month growth
plan. include customer acquisition
strategy, price and experiments, and
retention tactics. Show your reasoning.
With Deepthink on, Deep Seek spends
about 6 seconds thinking. It's breaking
down the problem, customer personas,
acquisition channels, pricing,
sensitivity, retention metrics. Then it
delivers a structured six-month plan
with specific action items for each
month. This is where the extended
reasoning really pays off. You're not
getting a generic answer. you're getting
a plan that actually considered
tradeoffs. Finally, let's test tool use.
This is where V3.2's new architecture
shines. I'm going to ask it to calculate
compound interest and show me the
formula step by step. Prompt, calculate
the compound interest on $10,000
invested at 5% annually for 10 years.
Show each step of the formula and the
final result. Deepseek doesn't just give
me the answer. It breaks down the
formula. explains what each variable
means, shows the calculation at each
step, then gives the final number in a
clean box. This is thank plus tool
execution in one flow. For anyone doing
research, analysis or learning, this is
huge. You're not just getting results.
You're seeing the work. Let's talk
honestly. Can Deepseek actually replace
Chat GPT or Gemini? The answer is it
depends. Here's where Deepseek wins.
Cost. Deepseek Chat is free, but API
usage is paid. Chat GPT Plus is $20 a
month. Claude Pro is 20. Google AI Pro
with Gemini is about $19.99 per month in
the US and around €21.99
per month in Germany. Varies by region
and taxes. If you're just using the chat
interface, Deep Seek cost you nothing.
Transparency. Deepseek often shows you
the visible thinking process. Chad GPT
and Gemini usually show the result and a
brief explanation without the full chain
of thought for learning debugging or
understanding why you got a certain
answer. Deepseek's transparency is a
massive advantage. Math and reasoning.
Deepseek R10528
scores 87.5%
on Me 2025 while GPT5 reports 94.6%
without tools. So GPT5 is ahead, but
Deep Seek is still very strong. If
you're doing data analysis, research, or
technical problem solving, Deep Seek
holds its own against most models. Open
weights. You can download the weights,
run locally, and use commercially
depending on the license. No vendor
lockin. For businesses worried about
data sovereignty, that's a big deal.
Here's where Chad GPT and Gemini still
win. Polish. Chad GBTs and Gemini's
interfaces are more refined, better
onboarding, smoother UX. Deep Seek feels
more utilitarian integrations. Chad GPT
has GTS with tools and actions, built-in
image generation, browsing, data
analysis. Gemini integrates with Google
Workspace. Deepseek is bare bones. You
get the model. That's it. Stability.
Deep See can show server busy messages
during peak demand. Paid incumbents
often feel more consistent, but no
service is immune to outages. Custom
instructions. Chat GPT has account level
custom instructions. Deepseek doesn't
yet offer the same kind of persistent
personalization in its chat UI, at least
not at the same level. Here's my honest
take. Can you replace chat GPT or Gemini
entirely? If you're doing writing,
research, analysis, and coding, yes,
absolutely. Deep Seek is good enough. If
you rely heavily on GPTs, need
guaranteed uptime, or want a more
polished experience, not yet. But for
80% of tasks, Deep Seek is shockingly
capable and it's free to use. Let me
show you a quick side by side. Same
prompt, both models. Prompt: Explain how
to improve website conversion rates in
five steps. Chad GPT gives a clean,
structured answer. Five steps, good
advice. Deepseek gives a clean,
structured answer. five steps. Also,
good advice. In this quick test, the
quality looks nearly identical. For most
everyday tasks, the performance gap is
minimal. The question is whether you
value the Chad GPT or Gemini ecosystem
or whether you just need a smart AI that
gets the job done. Now, let's talk about
prompting because the way you ask
matters. Deepseek isn't Chad GPT. It has
quirks and if you learn them, you'll get
way better results. Tip one, formatting
matters. If you're doing something
complex, structure your prompts with
clear sections. Deepseek responds
incredibly well to markdown style
formatting. Here's the template I use
for any non-trivial task. Task colon
what you want. Constraints colon any
limitations. Output colon how you want
the response formatted. For example,
task analyze this sales report.
Constraints focus on Q3 data only.
Ignore outliers above $10,000. output
bullet list with top three insights.
Each insight under 50 words. Watch how
clean that is. You've told the model
exactly what to do, what to avoid, and
what format you expect. No ambiguity, no
back and forth. There's also a second
format called think and answer. This
one's powerful for reasoning tasks. You
split your prompt into two sections
using tags. The think section is your
background context. The answer section
is your specific request. Here's an
example. Think I'm a freelance designer
launching a new service. My target
clients are small startups with budgets
under $5,000. I need a pricing strategy
that feels premium but accessible.
Answer: propose three pricing tiers with
specific dollar amounts and
justifications for each. The thank
section primes the model's reasoning.
The answer section narrows the output.
You get better, more tailored responses
this way. Tip two, be hyper specific.
Deep Seek needs detail way more than
Chad GBT. Don't just say write an email.
Say write a 200word professional email
to a client named Sarah explaining a
twoe project delay caused by supply
chain issues. Tone should be apologetic
but confident. Include a revised
timeline and next steps. See the
difference? You've given the model
length, tone, recipient context, reason
for the email and deliverables. Now, it
has everything it needs to nail it on
the first try. This applies to every
task. If you're asking for a comparison,
name both options and specify what
criteria matter to you. If you're asking
for a plan, state your constraints
upfront. Time, budget, team size,
whatever is relevant. Vague prompts give
you vague answers. Specific prompts give
you work you can actually use. Tip
three, use personas. Tell Deepse seeek
to act as someone. A career counselor, a
financial analyst, a creative director,
a systems engineer. This shapes the
style and knowledge it brings to the
response. Example, act as a senior
product manager. Review this feature
road map and identify the three biggest
risks to hitting our Q2 launch date. The
persona tells the model what lens to
use. You're not just asking for generic
feedback. You're asking for product
management feedback. The output changes.
I use this constantly for content work.
Act as a copywriter specializing in
landing pages. Rewrite this headline to
be more benefit driven and under 10
words. Boom. Focused output. No wasted
tokens. Tip four. Ask for multiple
approaches. Instead of asking how do I
solve this problem, ask give me three
ways to solve this problem with pros and
cons for each. This forces the model to
think deeper and consider trade-offs.
You're not getting one solution. You're
getting options. and options let you
make better decisions. Example, I need
to grow my email list by 5,000
subscribers in 3 months. Give me three
strategies. One, high effort, high
reward, one loweffort, lowreward, and
one experimental. Include estimated time
commitment and success probability for
each. That prompt gives you a strategic
breakdown, not just ideas, evaluated
ideas. Tip five, request selfch checks.
After getting an answer, ask what are
the potential risks with this approach?
Or what assumptions are you making here?
Deepseek's reasoning mode makes this
especially powerful. You'll literally
see it questioning its own logic. This
is how you catch blind spots before they
become problems. I do this on any high
stakes task. Legal advice, ask for
risks, financial projections, ask for
assumptions, strategic plan, ask what
could go wrong. The model will walk
through failure modes you might not have
considered. Tip six, placeholders for
templates. If you're writing something
reusable like an email template or a
sales script, use placeholders. Prompt,
write a cold email to potential clients,
use company name, your name, and product
name as placeholders. Then in the
response, every mention will use those
brackets. You just find and replace
later. This is huge for batch work. And
by the way, if you want 300 plus prompts
like this already written and ready to
copy paste, check out Prompt Lab inside
AMS Pro. I use them constantly.
Everything from freelance templates to
business automations saves me hours
every week. Tip seven, comparison
prompts. One thing I love doing is
setting up a pros cons list and then
asking the model to pick the best option
for a specific situation. Example,
compare working from home versus
co-working spaces. Give me five pros and
five cons for each. Then recommend which
one is best for a freelance software
developer with a tight budget and no
team. You get both the analysis and the
recommendation. And because DeepSeek
shows its reasoning, you can see exactly
why it picked one over the other. Tip
eight, iterate without fear. The first
response is rarely perfect. Follow up.
Make that more concise. Revise the
second paragraph to be less formal. Give
me a version that's more data driven.
The back and forth is how you dial in
exactly what you need. Don't treat the
first output as final. Treat it as a
starting point. Let me show you a quick
before and after. Bad prompt. How do I
market my product? Result from deepseek.
Vague, generic advice, social media,
email marketing, maybe try ads. Nothing
you couldn't Google in 5 seconds. Good
prompt. I'm launching a B2B SAS product
for small marketing teams. Budget is
$5,000 for the first quarter. What are
three lowcost customer acquisition
channels I should test and what metrics
should I track for each result from
deepseek specific actionable advice
LinkedIn organic outreach with reply
rate as the key metric content marketing
via SEO optimized blog posts track and
domain authority and organic traffic
partner referrals with conversion rate
and cost per acquisition clear next
steps for each that's the difference
specificity wins every time is this to
Chad GBT or Gemini Killer? No, it's an
alternative. For most people, it's good
enough to save $240 a year. The real
question is this. Do you actually need
Chat GPT's ecosystem or Gemini's Google
Workspace integration, or do you just
need a smart AI that writes well,
reasons clearly, and handles data? If
it's the latter, Deepseek delivers. My
recommendation, try it for one week. Use
it for your actual work, writing,
research, analysis, whatever you
normally throw at Cad GBT or Gemini. See
if you miss anything. I'm betting most
of you won't. And if you do need Chat
GBT for specific GPTs or Gemini for
Google integration, fine. But use
Deepseek for everything else. There's no
reason to pay for tasks a free model can
handle just as well. One more thing, the
fact that this model is open weight
matters. You can download it, you can
modify, you can run it on your own
hardware. If you've got this setup for
developers, researchers, and businesses
building AI products, this is a
gamecher. You're not locked into OpenAI
or Google. You've got options now. And
if you want more AI tools like this,
free, powerful, and actually useful,
subscribe. I test these every week and
break them down so you can start using
them right away. And if you are serious
about mastering AI beyond just one tool,
check out AI Master Pro in the
description below. That's where I keep
all my workflows, prompts, and training
in one place. And see you next time.