[Ken Shirriff] has been sharing a really low-level look at Intel’s Pentium (1993) processor. The Pentium’s architecture was highly innovative in many ways, and one of [Ken]’s most recent discoveries is that it contains a complex circuit — containing around 9,000 transistors — whose sole purpose is to multiply specifically by three. Why does such an apparently simple operation require such a complex circuit? And why this particular operation, and not something else?

Let’s back up a little to put this all into context. One of the feathers in the Pentium’s cap was its Floating Point Unit (FPU) which was capable of much faster floating point operations than any of its predecessors. [Ken] dove into reverse-engineering the FPU earlier this year and a close-up look at the Pentium’s silicon die shows that the FPU occupies a significant chunk of it. Of the FPU, nearly half is dedicated to performing multiplications and a comparatively small but quite significant section of that is specifically for multiplying a number by three. [Ken] calls it the x3 circuit.

Why does the multiplier section of the FPU in the Pentium processor have such specialized (and complex) functionality for such an apparently simple operation? It comes down to how the Pentium multiplies numbers.

Multiplying two 64-bit numbers is done in base-8 (octal), which ultimately requires fewer operations than doing so in base-2 (binary). Instead of handling each bit separately (as in binary multiplication), three bits of the multiplier get handled at a time, requiring fewer shifts and additions overall. But the downside is that multiplying by three must be handled as a special case.

[Ken] gives an excellent explanation of exactly how all that works (which is also an explanation of the radix-8 Booth’s algorithm) but it boils down to this: there are numerous shortcuts for multiplying numbers (multiplying by two is the same as shifting left by 1 bit, for example) but multiplying by three is the only one that doesn’t have a tidy shortcut. In addition, because the result of multiplying by three is involved in numerous other shortcuts (x5 is really x8 minus x3 for example) it must also be done very quickly to avoid dragging down those other operations. Straightforward binary multiplication is too slow. Hence the reason for giving it so much dedicated attention.

[Ken] goes into considerable detail on how exactly this is done, and it involves carry lookaheads as a key element to saving time. He also points out that this specific piece of functionality used more transistors than an entire Z80 microprocessor. And if that is not a wild enough idea for you, then how about the fact that the Z80 has a new OS available?

Lucid dreaming is the state of becoming aware one is dreaming while still being within the dream. To what end? That awareness may allow one to influence the dream itself, and the possibilities of that are obvious and compelling enough that plenty of clever and curious people have formed some sort of interest in this direction. Now there are some indications that VR might be a useful tool in helping people achieve lucid dreaming.

The research paper (Virtual reality training of lucid dreaming) is far from laying out a conclusive roadmap, but there’s enough there to make the case that VR is at least worth a look as a serious tool in the quest for lucid dreaming.

One method of using VR in this way hinges on the idea that engaging in immersive VR content can create mild dissociative experiences, and this can help guide and encourage users to perform “reality checks”. VR can help such reality checks become second nature (or at least more familiar and natural), which may help one to become aware of a dream state when it occurs.

Another method uses VR as a way to induce a mental state that is more conducive to lucid dreaming. As mentioned, engaging in immersive VR can induce mild dissociative experiences, so VR slowly guides one into a more receptive state before falling asleep. Since sleeping in VR is absolutely a thing, perhaps an enterprising hacker with a healthy curiosity in lucid dreaming might be inspired to experiment with combining them.

We’ve covered plenty of lucid dreaming hacks over the years and there’s even been serious effort at enabling communication from within a dreaming state. If you ask us, that’s something just begging to be combined with VR.

[Nicholas LaBonte] shows off a Cyberdeck Handheld that demonstrates just how good something can look when care and attention goes into the design and fabrication. He wanted to make something that blended cyberpunk and nautical aesthetics with a compact and elegant design, and we think he absolutely succeeded.

On the inside is a Raspberry Pi and an RTL-SDR. The back of the unit is machined from hardwood, and sports a bronze heat sink for the Raspberry Pi. The front has a prominent red PSP joystick for mouse input and a custom keyboard. The keyboard is especially interesting. On the inside it’s a custom PCB with tactile switches and a ATmega32U4 running QMK firmware — a popular choice for DIY keyboards — and presents to the host as a regular USB HID device.

How did he make those slick-looking keys? It’s actually a single plate that sits between the front panel and the switches themselves. [Nicholas] used a sheet of polymer with a faux-aluminum look to it and machined it down, leaving metal-looking keys with engraved symbols as tabs in a single panel. It looks really good, although [Nicholas] already has some ideas about improving it.

On the right side is the power button and charging port, and astute readers may spot that the power button is where a double-stack of USB ports would normally be on a Raspberry Pi 5. [Nicholas] removed the physical connectors, saving some space and connecting the USB ports internally to the keyboard and SDR.

As mentioned, [Nicholas] is already full of ideas for improvements. The bronze heat sink isn’t as effective as he’d like, the SDR could use some extra shielding, and the sounds the keyboard ends up making could use some work. Believe it or not, there’s still room to spare inside the unit and he’d maybe like to figure out a way to add a camera, GPS receiver, or maybe a 4G modem. We can’t wait! Get a good look for yourself in the video, embedded below.

Holography is about capturing 3D data from a scene, and being able to reconstruct that scene — preferably in high fidelity. Holography is not a new idea, but engaging in it is not exactly a point-and-shoot affair. One needs coherent light for a start, and it generally only gets touchier from there. But now researchers describe a new kind of holographic camera that can capture a scene better and faster than ever. How much better? The camera goes from scene capture to reconstructed output in under 30 milliseconds, and does it using plain old incoherent light.

The new camera is a two-part affair: acquisition, and calculation. Acquisition consists of a camera with a custom electrically-driven liquid lens design that captures a focal stack of a scene within 15 ms. The back end is a deep learning neural network system (FS-Net) which accepts the camera data and computes a high-fidelity RGB hologram of the scene in about 13 ms.  How good are the results? They beat other methods, and reconstruction of the scene using the data looks really, really good.

One might wonder what makes this different from, say, a 3D scene captured by a stereoscopic camera, or with an RGB depth camera (like the now-discontinued Intel RealSense). Those methods capture 2D imagery from a single perspective, combined with depth data to give an understanding of a scene’s physical layout.

Holography by contrast captures a scene’s wavefront information, which is to say it captures not just where light is coming from, but how it bends and interferes. This information can be used to optically reconstruct a scene in a way data from other sources cannot; for example allowing one to shift perspective and focus.

Being able to capture holographic data in such a way significantly lowers the bar for development and experimentation in holography — something that’s traditionally been tricky to pull off for the home gamer.

What does one do when frustrated at the lack of affordable, open source portable trackers? If you’re [OG-star-tech], you design your own and give it modular features that rival commercial offerings while you’re at it.

What’s a star tracker? It’s a method of determining position based on visible stars, but when it comes to astrophotography the term refers to a sort of hardware-assisted camera holder that helps one capture stable long-exposure images. This is done by moving the camera in such a way as to cancel out the effects of the Earth’s rotation. The result is long-exposure photographs without the stars smearing themselves across the image.

Interested? Learn more about the design by casting an eye over the bill of materials at the GitHub repository, browsing the 3D-printable parts, and maybe check out the assembly guide. If you like what you see, [OG-star-tech] says you should be able to build your own very affordably if you don’t mind 3D printing parts in ASA or ABS. Prefer to buy a kit or an assembled unit? [OG-star-tech] offers them for sale.

Frustration with commercial offerings (or lack thereof) is a powerful motive to design something or contribute to an existing project, and if it leads to more people enjoying taking photos of the night sky and all the wonderful things in it, so much the better.

[Martin] of [Wintergatan] is on a quest to create the ultimate human-powered, modern marble music machine. His fearless mechanical exploration and engineering work, combined with considerable musical talent, has been an ongoing delight as he continually refines his designs. We’d like to highlight this older video in which he demonstrates how to dynamically regulate the speed of a human-cranked music machine by taking inspiration from gramophones: he uses a flyball governor (or centrifugal governor).

These devices are a type of mechanical feedback system that was invented back in the 17th century but really took off once applied to steam engines. Here’s how they work: weights are connected to a shaft with a hinged assembly. The faster the shaft spins, the more the weights move outward due to centrifugal force. This movement is used to trigger some regulatory action, creating a feedback loop. In a steam engine, the regulator adjusts a valve which keeps the engine within a certain speed range. In a gramophone it works a wee bit differently, and this is the system [Wintergatan] uses.

To help keep the speed of his music machine within a certain narrow range, instead of turning a valve the flyball governor moves a large disk brake. The faster the shaft spins, the harder the brake is applied. Watch it in action in the video (embedded below) which shows [Wintergatan]’s prototype, demonstrating how effective it is.

[Wintergatan]’s marble machine started out great and has only gotten better over the years, with [Martin] tirelessly documenting his improvements on everything. After all, when every note is the product of multiple physical processes that must synchronize flawlessly, it makes sense to spend time doing things like designing the best method of dropping balls.

One final note: if you are the type of person to find yourself interested and engaged by these sorts of systems and their relation to obtaining better results and tighter tolerances, we have a great book recommendation for you.

[JanTec Engineering] was fascinated by the idea of using a 3D printer’s hot end to inject voids and channels in the infill with molten plastic, leading to stronger prints without the need to insert hardware or anything else. Inspiration came from two similar ideas: z-pinning which creates hollow vertical channels that act as reinforcements when filled with molten plastic by the hot end, and VoxelFill (patented by AIM3D) which does the same, but with cavities that are not uniform for better strength in different directions. Craving details? You can read the paper on z-pinning, and watch VoxelFill in (simulated) action or browse the VoxelFill patent.

With a prominent disclaimer that his independent experiments are not a copy of VoxelFill nor are they performing or implying patent infringement, [JanTec] goes on to use a lot of custom G-code (and suffers many messy failures) to perform some experiments and share what he learned.

One big finding is that one can’t simply have an empty cylinder inside the print and expect to fill it all up in one go. Molten plastic begins to cool immediately after leaving a 3D printer’s nozzle, and won’t make it very far down a deep hole before it cools and hardens. One needs to fill a cavity periodically rather than all in one go. And it’s better to fill it from the bottom-up rather than from the top-down.

He got better performance by modifying his 3D printer’s hot end with an airbrush nozzle, which gave about 4 mm of extra length to work with. This extra long nozzle could reach down further into cavities, and fill them from the bottom-up for better results. Performing the infill injection at higher temperatures helped fill the cavities more fully, as well.

Another thing learned is that dumping a lot of molten plastic into a 3D print risks deforming the print because the injected infill brings a lot of heat with it. This can be mitigated by printing the object with more perimeters and a denser infill so that there’s more mass to deal with the added heat, but it’s still a bit of a trouble point.

[JanTec] put his testing hardware to use and found that parts with infill injection were noticeably more impact resistant than without. But when it came to stiffness, an infill injected part resisted bending only a little better than a part without, probably because the test part is very short and the filled cavities can’t really shine in that configuration.

These are just preliminary results, but got him thinking there are maybe there are possibilities with injecting materials other than the one being used to print the object itself. Would a part resist bending more if it were infill injected with carbon-fibre filament? We hope he does some follow-up experiments; we’d love to see the results.

[James Sharman] designed and built his own 8-bit computer from scratch using TTL logic chips, including a VGA adapter, and you can watch it run a glorious rotating cube demo in the video below.

The rotating cube is the product of roughly 3,500 lines of custom assembly code and looks fantastic, running at 30 frames per second with shading effects from multiple light sources. Great results considering the computing power of his system is roughly on par with vintage 8-bit home computers, and the graphics capabilities are limited. [James]’s computer uses a tile map instead of a frame buffer, so getting 3D content rendered was a challenge.

The video is about 20 seconds of demo followed by a detailed technical discussion on how exactly one implements everything required for a 3D cube, from basic math to optimization. If a deep dive into that sort of thing is up your alley, give it a watch!

We’ve featured [James]’ fascinating work on his homebrew computer before. Here’s more detail on his custom VGA adapter, and his best shot at making it (kinda) run DOOM.

One can 3D print with conductive filament, and therefore plausibly create passive components like resistors. But what about active components, which typically require semiconductors? Researchers at MIT demonstrate working concepts for a resettable fuse and logic gates, completely 3D printed and semiconductor-free.

Now just to be absolutely clear — these are still just proofs of concept. To say they are big and perform poorly compared to their semiconductor equivalents would be an understatement. But they do work, and they are 100% 3D printed active electronic components, using commercially-available filament.

How does one make a working resettable fuse and transistor out of such stuff? By harnessing thermal expansion, essentially.

The conductive filament the researchers used is Electrifi by Multi3D, which is PLA combined with copper micro-particles. A segment printed in this filament is normally very conductive due to the densely-packed particles, but as temperature increases (beginning around 40° C) the polymer begins to soften and undergoes thermal expansion. This expansion separates the copper particles, causing a dramatic increase in electrical resistance as electrical pathways are disrupted. That’s pretty neat, but what really ties it together is that this behavior is self-resetting, and reversible. As long as the PLA isn’t straight up melted (that is to say, avoids going over about 150° C) then as the material cools it contracts and restores the conductive pathways to their original low-resistance state. Neat!

So where does the heat required come from? Simply passing enough current through the junction will do the job. By carefully controlling the size and shape of traces (something even hobbyist filament-based 3D printers are very good at) this effect can be made predictable and repeatable.

The simpler of the two test components uses the resistance spike as a self-resetting fuse. The printed component is designed such that current above a threshold triggers a surge in resistance, preventing damage to some theoretical circuitry downstream. As long as the component is not destroyed by heating it to the point that it melts, it self-resets as it cools.

The transistor is a bit more interesting. By designing two paths so that they intersect each other, one can be used as a control path and the other as a signal path. Applying a voltage to the control path electrically controls the resistance of the signal path, effectively acting as a transistor. Researchers combined these basic transistors into NOT, AND, and OR gates. One is shown here.

This whole system is scalable, low-cost, and highly accessible to just about anyone with some basic equipment. Of course, it has some drawbacks. The switching speed is slow (seconds rather than nanoseconds) and being thermally-driven means power consumption is high. Still, it’s pretty nifty stuff. Check out the research paper for all the nitty-gritty details.

We’ve seen 3D printed triboelectric generators so it’s pretty exciting to now see printed active electronic components. Maybe someday they can be combined?

OpenSCAD has a lot of fans around these parts — if you’re unaware, it’s essentially a code-based way of designing 3D models. Instead of drawing them up in a CAD program, one writes a script that defines the required geometry. All that is made a little easier with the Belfry OpenSCAD Library (BOSL2).

BOSL2 has an extensive library of base shapes, advanced functions for manipulating models, and some really nifty tools for creating attachment points on parts and aligning components with one another. If that sounds handy for designing useful objects, you’re in for even more of a treat when you see their functions for gears, hinges, screws, and more.

There’s even one that covers bottle necks and caps. (Those are all standardized by the way, so it’s never been easier to interface to existing bottles or caps in a project.)

OpenSCAD really is very versatile software. It powers useful tools like this screw, washer, and nut generator as well as having more unusual applications like a procedural terrain generator. It’s free, so if you’ve never looked into it, check it out!

What would it be like to have to design and build a ventilator, suitable for clinical use, in ten days? One that could be built entirely from locally-sourced parts, and kept oxygen waste to a minimum? This is the challenge [John Dingley] and many others faced at the start of COVID-19 pandemic when very little was known for certain.

Back then it was not even known if a vaccine was possible, or how bad it would ultimately get. But it was known that hospitalized patients could not breathe without a ventilator, and based on projections it was possible that the UK as a whole could need as many as 30,000 ventilators within eight weeks. In this worst-case scenario the only option would be to build them locally, and towards that end groups were approached to design and build a ventilator, suitable for clinical use, in just ten days.

[John] decided to create a documentary called Breathe For Me: Building Ventilators for a COVID Apocalypse, not just to tell the stories of his group and others, but also as a snapshot of what things were like at that time. In short it was challenging, exhausting, occasionally frustrating, but also rewarding to be able to actually deliver a workable solution.

In the end, building tens of thousands of ventilators locally wasn’t required. But [John] felt that the whole experience was a pretty unique situation and a remarkable engineering challenge for him, his team, and many others. He decided to do what he could to document it, a task he approached with a typical hacker spirit: by watching and reading tutorials on everything from conducting and filming interviews to how to use editing software before deciding to just roll up his sleeves and go for it.

We’re very glad he did, and the effort reminds us somewhat of the book IGNITION! which aimed to record a history of technical development that would otherwise have simply disappeared from living memory.

You can watch Breathe for Me just below the page break, and there’s additional information about the film if you’d like to know a bit more. And if you are thinking the name [John Dingley] sounds familiar, that’s probably because we have featured his work — mainly on self-balancing personal electric vehicles — quite a few times in the past.

Disk space is allocated in clusters of a certain size. When a file is written to disk and the file size is smaller than the cluster(s) allocated for it, there is an unused portion of varying size between the end of the file’s data and the end of the allocated clusters. This unused space is the slack space, it’s perfectly normal, and [Zachary Parish] had an idea to write a tool to hide data in it.

[Zachary]’s tool is in Python and can map available slack space and perform read and write operations on it, treating the disparate locations as a single unified whole in which to store arbitrary files. A little tar and gzip even helps makes things more efficient in the process.

There’s a whole demo implemented on Linux using a usb drive with some decoy files to maximize the slack space, and you can watch it in action in the video embedded below. It’s certainly more practical than hiding data in a podcast!

Note that this is just a demo of the concept. The approach does have potential for handling secret data, but [Zachary] points out that there are — from a serious data forensics point of view– a number of shortcomings in its current form. For example, the way the tool currently structures and handles data makes it quite obvious that something is going on in the slack space.

[Zachary] created this a few years ago and has some ideas about how to address those shortcomings and evolve the tool, so if you have ideas of your own or just want to try it out, the slack_hider GitHub repository is where you want to go.

Bleach is a handy way to mark fabrics, and it turns out that combining bleach with a 3D-printed design is an awfully quick-working and effective way to stamp a design onto a shirt.

While conceptually simple, the details make the difference. Spraying bleach onto the stamp surface helps get even coverage, and having the stamp facing “up” and lowering the shirt onto the stamp helps prevent bleach from running where it shouldn’t. Prompt application of hot air with a heat gun (followed by neutralizing or flushing any remaining bleach by rinsing in plenty of cold water) helps keep the edges of the design clean and sharp.

We wondered if combining techniques with some of the tips on how to 3D print ink stamps would yield even better results. For instance, we notice the PLA stamp (used to make the design in the images here) produces sharp lines with a slightly “eroded” look overall. This is very much like the result of inking with a stamp printed in PLA. With a stamp printed in flex filament, inking gives much more even results, and we suspect the same might be true for bleach.

Of course, don’t forget that it’s possible to 3D print directly onto fabric if you want your designs to be a little more controlled (and possibly in multiple colors). Or, try silkscreening. Who knew there were so many options for putting designs onto shirts? If you try it out and learn anything, let us know by sending in a tip!

Bear with us for a moment for a little background. The Rideau Canal Skateway in Ottawa is the world’s largest natural skating rink, providing nearly 8 km of pristine ice surface during the winter. But maintaining such a large ice surface is a challenge. A regular Zamboni can’t do it; the job is just too big. So the solution is a custom machine called the Froster, conceived by Robert Taillefer and built by Sylvain Fredette.

A patent was filed in 2010, granted by the Canadian Intellectual Property Office, and later lost because important notifications started going to an apparently unchecked spam folder. The annual fee went unpaid, numerous emails went unanswered, an expiry date came and went, and that was that.

It’s true that emailed reminders (the agreed-upon — and only — method of contact) going unnoticed to spam was what caused Robert to not take any action until it was too late. We’d all agree that digital assistants in general need to get smarter, and that includes being better at informing the user about automatically-handled things like spam.

But what truly cost Robert Taillefer his patent was having a single point of failure for something very, very important. The lack of any sort of backup method of communication in case of failure or problem meant that this sad experience was, in a way, a disaster just waiting to happen. At least that’s how the Federal Court saw it when he took his complaint to them, and that’s how they continued to see it when he appealed the decision.

If you’ve never heard of the Rideau Canal Skateway or would like to see the Froster in action, check out this short video from the National Capital Commission of Canada, embedded just under the page break.

Do you make things, and have you got almost ten minutes to spare? If not, make the time because this video by [PrintLab] is chock-full of healthy and practical design tips. It’s about effective design of Assistive Technology, but the design concepts extend far beyond that scope.

It’s about making things that are not just functional tools, but objects that are genuinely desirable and meaningful to people’s lives. There are going to be constraints, but constraints aren’t limits on creativity. Heck, some of the best devices are fantastic in their simplicity, like this magnetic spoon.

One item that is particularly applicable in our community is something our own [Jenny List] has talked about: don’t fall into the engineer-saviour trap. The video makes a similar point in that it’s easy and natural to jump straight into your own ideas, but it’s critical not to make assumptions. What works in one’s head may not work in someone’s actual life. The best solutions start with a solid and thorough understanding of an issue, the constraints, and details of people’s real lives.

Another very good point is that designs don’t spring fully-formed from a workbench, so prototype freely using cardboard, models, 3D printing, or whatever else makes sense to you. Don’t be stingy with your prototyping! As long as you’re learning something each time, you’re on the right path.

And when a design is complete? It has the potential to help others, so share it! But sharing and opening your design isn’t just about putting the files online. It’s also about making it as easy as possible for others to recreate, integrate, or modify your work for their own needs. This may mean making clear documentation or guides, optimizing your design for ease of editing, and sharing the rationale behind your design choices to help others can build on your work effectively.

The whole video is excellent, and it’s embedded here just under the page break. Does designing assistive technology appeal to you? If so, then you may be interested in the Make:able challenge which challenges people to design and make a 3D printable product (or prototype) that improves the day-to-day life of someone with a disability, or the elderly. Be bold! You might truly help someone’s life.

This technique shared by [Andy Kong] is for 3D printed lenses, but would probably be worth a shot for any resin prints that need to be made nice and clear. The link to his post on X is here, but we’ll summarize below.

It’s entirely possible to print lenses on a resin printer, but some amount of polishing is inevitable because an SLA print still has layer lines, however small. We have seen ways to minimize the work involved to get a usable lens, but when it comes right down to it the printing process creates tiny (but inevitable) surface imperfections that have to be dealt with, one way or another.

One technique involves applying a thin layer of liquid resin to the surface of the printed lens, then curing it. This isn’t a complete solution because getting an even distribution of resin over the surface can be a challenge. [Andy] has refined this technique to make it ridiculously simple, and here’s how it works.

After printing the lens, place a drop of liquid resin on the lens surface and stretch some cling wrap over the lens. The cling wrap conforms to the shape and curve of the lens while trapping a super thin layer of liquid resin between the cling wrap film and the lens surface. One then cures the resin while holding the cling film taut. After curing, [Andy] says the film peels right off, leaving an ultra-smooth surface behind. No tedious polishing required!

But what about the flat back of the lens? [Andy] suggests that instead of using cling film (which is better at conforming to a curved surface) simply use a drop of resin in a similar way to bond the flat side of the lens to a smooth piece of glass. Or bond the backs of two lenses together to make a duplex lens. This technique opens quite a few possibilities!

Even if one isn’t 3D printing optical lenses, we suspect this technique might be applicable to making crystal-clear 3D prints with a little less effort than would otherwise be needed.

Keep it in mind, and if you find success (or failure!) let us know on the tips line because we absolutely want to hear about it.

The DeepSeek large language models (LLM) have been making headlines lately, and for more than one reason. IEEE Spectrum has an article that sums everything up very nicely.

We shared the way DeepSeek made a splash when it came onto the AI scene not long ago, and this is a good opportunity to go into a few more details of why this has been such a big deal.

For one thing, DeepSeek (there’s actually two flavors, -V3 and -R1, more on them in a moment) punches well above its weight. DeepSeek is the product of an innovative development process, and freely available to use or modify. It is also indirectly highlighting the way companies in this space like to label their LLM offerings as “open” or “free”, but stop well short of actually making them open source.

The DeepSeek-V3 LLM was developed in China and reportedly cost less than 6 million USD to train. This was possible thanks to developing DualPipe, a highly optimized and scalable method of training the system despite limitations due to export restrictions on Nvidia hardware. Details are in the technical paper for DeepSeek-V3.

There’s also DeepSeek-R1, a chain-of-thought “reasoning” model which handily provides its thought process enclosed within easily-parsed <think> and </think> pseudo-tags that are included in its responses. A model like this takes an iterative step-by-step approach to formulating responses, and benefits from prompts that provide a clear goal the LLM can aim for. The way DeepSeek-R1 was created was itself novel. Its training started with supervised fine-tuning (SFT) which is a human-led, intensive process as a “cold start” which eventually handed off to a more automated reinforcement learning (RL) process with a rules-based reward system. The result avoided problems that come from relying too much on RL, while minimizing the human effort of SFT. Technical details on the process of training DeepSeek-R1 are here.

DeepSeek-V3 and -R1 are freely available in the sense that one can access the full-powered models online or via an app, or download distilled models for local use on more limited hardware. It is free and open as in accessible, but not open source because not everything needed to replicate the work is actually released. Like with most LLMs, the training data and actual training code used are not available.

What is released and making waves of its own are the technical details of how researchers produced what they did, and that means there are efforts to try to make an actually open source version. Keep an eye out for Open-R1!