Results for
I've talked about running local Large Language Models a couple of times on The MATLAB Blog but always had to settle for small models because of the tiny amount of memory on my GPU -- 6GB to be precise! Running much larger, more capable models meant requireing expensive, sever-class GPUs on HPC or cloud instances and I never had enough budget to do it.
Until now!
NVIDIA's DGX Spark is a small desktop machine that doesn't cost the earth. Indeed, several of us at MathWorks have one now although 'mine' (pictured above sporting a MATLAB sticker) is actually shared with a few other people and lives on a desk in Natick, USA while I'm in the UK.
The DGX Spark has 128GB of memory available to the GPU which means that I can run a MUCH larger language model. So, I installed a 120 Billion parameter model on it: gpt-oss:120b. More than an order of magnitude bigger than any local model I had played with before.
The next step was to connect to it from MATLAB running on my laptop.
The result is a *completely private* MATLAB + AI workflow that several of us have been playing with.
In my latest article, I show you how to set everything up: The LLM running on the DGX Spark connected to MATLAB running on my MacBook Pro. https://blogs.mathworks.com/matlab/2026/01/05/running-large-language-models-on-the-nvidia-dgx-spark-and-connecting-to-them-in-matlab/
I thought it would be fun to vibe code Christmas trees.
Model: Claude Sonnet4.5
Prompt: Write a MATLAB script that draws a Christmas tree. There should be decorations and presents too.
Result:
christmasTree()
function christmasTree()
% Christmas Tree with Decorations and Presents
% Creates a festive scene with a decorated tree and gift boxes
figure('Color', 'w', 'Position', [100, 100, 800, 900]);
hold on;
axis equal;
axis off;
xlim([-3, 3]);
ylim([0, 10]);
% Draw the tree (three layered triangular sections)
drawTreeSection(0, 2, 2.5, 2.5, [0.1, 0.5, 0.1]);
drawTreeSection(0, 3.5, 2, 2, [0.15, 0.55, 0.15]);
drawTreeSection(0, 4.8, 1.5, 1.8, [0.2, 0.6, 0.2]);
% Tree trunk
rectangle('Position', [-0.3, 0, 0.6, 2], 'FaceColor', [0.4, 0.2, 0.1], ...
'EdgeColor', 'none');
% Add decorations
addOrnaments();
addTinsel();
% Star on top
drawStar(0, 7.2, 0.4, [1, 0.84, 0]);
% Add presents under the tree
drawPresent(-1.8, 0, 0.6, 0.5, [0.8, 0.1, 0.1], [1, 0.84, 0]);
drawPresent(-0.8, 0, 0.5, 0.6, [0.1, 0.3, 0.8], [0.9, 0.9, 0.9]);
drawPresent(1.2, 0, 0.7, 0.4, [0.7, 0.5, 0.9], [0.2, 0.8, 0.2]);
drawPresent(0.2, 0, 0.4, 0.5, [1, 0.7, 0.1], [0.8, 0.1, 0.1]);
% Add some snow on the ground
scatter(randn(1, 100)*2.5, rand(1, 100)*0.3, 20, 'w', 'filled', ...
'MarkerEdgeColor', [0.8, 0.8, 0.8]);
title('Merry Christmas!', 'FontSize', 24, 'Color', [0.8, 0.1, 0.1], ...
'FontWeight', 'bold');
hold off;
end
function drawTreeSection(x, y, width, height, color)
% Draw a triangular section of the tree
xPoints = [x - width, x + width, x];
yPoints = [y, y, y + height];
fill(xPoints, yPoints, color, 'EdgeColor', [0.05, 0.3, 0.05], ...
'LineWidth', 1.5);
end
function addOrnaments()
% Add colorful baubles to the tree
ornamentColors = [
1, 0, 0; % Red
0, 0, 1; % Blue
1, 0.84, 0; % Gold
0.9, 0.1, 0.5;% Pink
0.5, 0, 0.5 % Purple
];
% Ornament positions (x, y coordinates)
positions = [
-1.5, 2.5; 1.5, 2.5; 0, 3.2;
-1.2, 3.8; 1.2, 3.8; -0.6, 4.2; 0.6, 4.2;
-0.8, 5.2; 0.8, 5.2; 0, 5.8; -0.4, 6.3; 0.4, 6.3
];
for i = 1:size(positions, 1)
colorIdx = mod(i-1, size(ornamentColors, 1)) + 1;
x = positions(i, 1);
y = positions(i, 2);
% Draw bauble
theta = linspace(0, 2*pi, 50);
radius = 0.15;
xCirc = x + radius * cos(theta);
yCirc = y + radius * sin(theta);
fill(xCirc, yCirc, ornamentColors(colorIdx, :), ...
'EdgeColor', ornamentColors(colorIdx, :)*0.7, 'LineWidth', 1.5);
% Add highlight for shine
highlightX = x - radius*0.4 + radius*0.3*cos(theta(1:15));
highlightY = y + radius*0.4 + radius*0.3*sin(theta(1:15));
fill(highlightX, highlightY, [1, 1, 1], 'EdgeColor', 'none', ...
'FaceAlpha', 0.6);
end
end
function addTinsel()
% Add sparkly tinsel effect
tinselX = [-1.8, -1.2, -0.5, 0.2, 0.9, 1.6];
tinselY = [2.8, 3.5, 4.0, 4.5, 5.0, 5.5];
for i = 1:length(tinselX)
plot(tinselX(i) + randn(1, 5)*0.1, tinselY(i) + rand(1, 5)*0.2, ...
'*', 'Color', [1, 0.84, 0], 'MarkerSize', 8, 'LineWidth', 1.5);
end
end
function drawStar(x, y, size, color)
% Draw a 5-pointed star on top of the tree
angles = linspace(0, 2*pi, 11);
radii = repmat([size, size*0.4], 1, 5);
radii = [radii, radii(1)];
xStar = x + radii .* cos(angles - pi/2);
yStar = y + radii .* sin(angles - pi/2);
fill(xStar, yStar, color, 'EdgeColor', [0.8, 0.6, 0], 'LineWidth', 2);
% Add sparkle
plot(x, y, '*', 'Color', [1, 1, 1], 'MarkerSize', 15, 'LineWidth', 2);
end
function drawPresent(x, y, width, height, boxColor, ribbonColor)
% Draw a gift box with ribbon and bow
% Box
rectangle('Position', [x, y, width, height], 'FaceColor', boxColor, ...
'EdgeColor', boxColor*0.7, 'LineWidth', 2);
% Vertical ribbon
ribbonWidth = width * 0.15;
rectangle('Position', [x + width/2 - ribbonWidth/2, y, ribbonWidth, height], ...
'FaceColor', ribbonColor, 'EdgeColor', 'none');
% Horizontal ribbon
ribbonHeight = height * 0.15;
rectangle('Position', [x, y + height/2 - ribbonHeight/2, width, ribbonHeight], ...
'FaceColor', ribbonColor, 'EdgeColor', 'none');
% Bow on top
bowX = x + width/2;
bowY = y + height;
bowSize = width * 0.2;
% Left loop
theta = linspace(0, pi, 30);
fill(bowX - bowSize*0.3 + bowSize*0.5*cos(theta), ...
bowY + bowSize*0.5*sin(theta), ribbonColor, 'EdgeColor', 'none');
% Right loop
fill(bowX + bowSize*0.3 + bowSize*0.5*cos(theta), ...
bowY + bowSize*0.5*sin(theta), ribbonColor, 'EdgeColor', 'none');
% Center knot
theta = linspace(0, 2*pi, 30);
fill(bowX + bowSize*0.25*cos(theta), bowY + bowSize*0.25*sin(theta), ...
ribbonColor*0.8, 'EdgeColor', 'none');
end
Large Language Models (LLMs) with MATLAB was updated again today to support the newly released OpenAI models GPT-5, GPT-5 mini, GPT-5 nano, GPT-5 chat, o3, and o4-mini. When you create an openAIChat object, set the ModelName name-value argument to "gpt-5", "gpt-5-mini", "gpt-5-nano", "gpt-5-chat-latest", "o4-mini", or "o3".
This is version 4.4.0 of this free MATLAB add-on that lets you interact with LLMs on MATLAB. The release notes are at Release v4.4.0: Support for GPT-5, o3, o4-mini · matlab-deep-learning/llms-with-matlab
Imagine you are developing a new toolbox for MATLAB. You have a folder full of a few .m files defining a bunch of functions and you are thinking 'This would be useful for others, I'm going to make it available to the world'
What process would you go through? What's the first thing you'd do?
I have my own opinions but don't want to pollute the start of the conversation :)
You've probably heard about the DeepSeek AI models by now. Did you know you can run them on your own machine (assuming its powerful enough) and interact with them on MATLAB?
In my latest blog post, I install and run one of the smaller models and start playing with it using MATLAB.
Larger models wouldn't be any different to use assuming you have a big enough machine...and for the largest models you'll need a HUGE machine!
Even tiny models, like the 1.5 billion parameter one I demonstrate in the blog post, can be used to demonstrate and teach things about LLM-based technologies.
Have a play. Let me know what you think.
When I want to understand a problem, I'll often use different sources. I'll read different textbooks, blog posts, research papers and ask the same question to different people. The differences in the solutions are almost always illuminating.
I feel the same way about AIs. Sometimes, I don't want to ask *THE* AI...I want to ask a bunch of them. They'll have different strengths and weaknesses..different personalities if you want to think of it that way.
I've been playing with the AI chat arena and there really is a lot of difference between the answers returned by different models. https://lmarena.ai/?arena
I think it would be great if the MATLAB Chat playgroundwere to allow the user to change which AI they were talking with.
What does everyone else think?
I recently wrote about the new ODE solution framework in MATLAB over the The MATLAB Blog The new solution framework for Ordinary Differential Equations (ODEs) in MATLAB R2023b » The MATLAB Blog - MATLAB & Simulink (mathworks.com)
This was a very popular post at the time - many thousands of views. Clearly everyone cares about ODEs in MATLAB.
This made me wonder. If you could wave a magic wand, what ODE functionality would you have next and why?
Every day, thousands of people ask questions on MATLAB Answers and many of these are about their code. Questions such as “How can I make this faster?”, “Why do I get this error message?” or “Why don’t I get the answer I expect?”. There’s often one crucial thing missing though – the code in question!
Most of the people who answer questions on MATLAB Answers are volunteers from the community. They are answering your questions for fun, to learn more about MATLAB or just because they like to be helpful. This is even true for people such as me who are MathWorks members of staff. It’s not part of my role to patrol the community, looking where I can help out. I do it because I like to do it.
Make it easier to help me help you.
Imagine you’re a volunteer, looking for something interesting to answer. What kind of questions are you more likely to dig into and help an anonymous stranger figure out?
In my case, I almost always focus on problems that I can easily reproduce. I rarely know the answer to any question off the top of my head and so what I like to do is start off with the problem you are facing and use the various tools available to me such as the profiler or debugger to figure it out. This is the fun of it all for me – I almost always learn something by doing this and you get helped out as a side effect!
The easier I can reproduce your issue, the more likely I am to get started. If I can’t reproduce anything and the answer isn’t immediately obvious to me I’ll just move onto the next question. One example that demonstrates this perfectly is a case where someone’s MATLAB code was running too slowly. All of the code was available so I could run it on my machine, profile it and provide a speed-up of almost 150x.
It's not always feasible or desirable to post all of your code in which case you need to come up with a minimal, reproducible example. What’s the smallest amount of code and data you can post that I can run on my machine and see what you see? This may be more work for you but it will greatly increase your chances of receiving an answer to your question.
