News

A new NYT Strands puzzle appears at midnight each day for your time zone – which means that some people are always playing 'today's game' while others are playing 'yesterday's'. If you're looking for Friday's puzzle instead then click here: NYT Strands hints and answers for Friday, September 11 (game #922).

Strands is the NYT's latest word game after the likes of Wordle, Spelling Bee and Connections – and it's great fun. It can be difficult, though, so read on for my Strands hints.

Want more word-based fun? Then check out my NYT Connections today and Quordle today pages for hints and answers for those games, and Marc's Wordle today page for the original viral word game.

SPOILER WARNING: Information about NYT Strands today is below, so don't read on if you don't want to know the answers.

• Today's NYT Strands theme is… Don’t obsess

Play any of these words to unlock the in-game hints system.

• EACH
• STEED
• BEAR
• GONE
• GUNGE
• TRIFLE
• LIFE

• Spangram has 7 letters

• First side: left, 4th row

• Last side: right, 5th row

Right, the answers are below, so DO NOT SCROLL ANY FURTHER IF YOU DON'T WANT TO SEE THEM.

The answers to today's Strands, game #923, are…

• LIBERATE
• FREE
• RELEASE
• DETACH
• UNLOOSE
• SPANGRAM: LETITGO

• My rating: Hard
• My score: 1 hint

I didn’t understand today’s theme at first and even after spotting LIBERATE I still needed a hint (FREE) before I finally got going.

I first heard of the phrase LETITGO in the 1990s, when a Buddhist rapper explained the concept to me and how it had helped them deal with fame and pressure.

Initially, I’d thought it was about possessions rather than feelings, but it’s a philosophy I adhere to in the main. That said, in my opinion, certain people in the world deserve all the negative energy we can fire in their direction — letting it go would just be letting them off the hook.

• BRIDGE
• BARGE
• SHIP
• DOCK
• PIER
• LIGHTHOUSE
• SPANGRAM: ONTHEWATERFRONT

Strands is the NYT's not-so-new-any-more word game, following Wordle and Connections. It's now a fully fledged member of the NYT's games stable that has been running for a year and which can be played on the NYT Games site on desktop or mobile.

I've got a full guide to how to play NYT Strands, complete with tips for solving it, so check that out if you're struggling to beat it each day.

• Ford CEO hits back at US government's China claims
• Ford’s new Fathom electric pick-up uses technology licensed from CATL
• The affordable Fathom is part of Ford’s new Universal EV platform

Ford has hit back at criticisms from US Transportation Secretary Sean Duffy that the automaker is becoming overly reliant on Chinese car companies, with Chief Executive Jim Farley telling The Wall Street Journal that these are “basic misunderstandings, mistruths, whatever words you want to use.”

The global automaker, which is winding up production of its current F-150 Lightning, is preparing a smaller, more affordable electric pick-up for the masses.

Dubbed Fathom, it will start at $28,350 (around £20,970 / AU$39,540) before destination fees and is the result of work by Ford’s “skunkworks” team, which was assembled to find ways to innovate in the EV space while keeping costs down and competing with increasingly capable Chinese electric vehicles.

However, the Transportation Secretary accused Ford of “actively intertwining its future with Chinese state-backed enterprises,” including a plant that licenses technology from Chinese battery giant CATL, according to the WSJ.

Ford announced in 2023 that it would begin manufacturing its own lithium iron phosphate (LFP) batteries at the newly formed BlueOval Battery Park Michigan in Marshall.

The company subsequently confirmed that the plant would use technology and expertise licensed from China-based CATL, but Ford would own and operate the facility and assemble the batteries in the US.

Production of the LFP battery cells destined for Ford’s affordable electric pick-up is now ramping up, with Ford saying it is on track to ship batteries in 2026. Farley is adamant that the arrangement does not amount to handing control of Ford’s future to China.

“We’re not enabling the Chinese to come here,” Farley told the WSJ. “Actually, you could argue the opposite is the case.”

In a statement published by Ford HQ on its media site, the company said that it is the "most American automaker".

"We produce more vehicles in the US, employ more hourly manufacturing workers and export more vehicles from the US than any other automaker", it added.

Despite a retirement in its current form, the F-150 Lightning itself isn't being killed off entirely, either. Ford says production of the current version will end this year, with a next-generation F-150 Lightning EREV planned to offer more than 700 miles of combined range using a battery and gasoline-powered generator.

Ford’s CEO has called the creation of its new Universal EV platform, which it hopes will underpin a whole host of new EVs, the company’s “Model T moment,” stating that the platform will help bring electric trucks and vehicles to customers at a more affordable price.

Full details of the Fathom are still to be revealed, but we do know that it will cost less than $30,000 before destination charges when it goes on sale, with a starting MSRP of $28,350 (about £20,970 / AU$39,540). It will use one of Ford’s new LFP battery packs and feature a NACS charging port.

It will also pack bidirectional charging, allowing owners to power tools and other accessories from the batteries, and every Fathom will have the hardware required for Ford’s next-generation BlueCruise hands-free highway driving system.

Four doors, five passengers, a generous rear bed and a front trunk for extra storage will all come as standard, while Ford is targeting a midsize footprint with more passenger space than a Toyota RAV4.

Ford hopes that its electric trucks will finally hit home with a mass-market audience, thanks to the technology on offer and extremely competitive pricing.

Although it will count the Jeff Bezos-backed Slate truck as a key competitor, that pick-up is priced from less than $25,000 (about £18,490 / AU$34,870) before destination charges.

• RAND report claims the civilian space contractor workforce in the US can not support rebuilding lost satellites within 12 to 24 months of a war with either Russia or China
• The benchmark timeline is based on RAND's own assumptions and calculations surrounding a protracted war with either regional superpower
• The report highlights labor gaps that are as dire as 19:1 in certain cases when it comes to engineering replacements

A new report has warned the United States might struggle to replace satellites lost in a space war because it lacks the requisite workers.

The report, "Breaking Glass, Missing Hands", was sponsored by the Vice Chief of Space Operations and conducted by RAND Project AIR FORCE, the Department of the Air Force's own federally funded research center, as part of a fiscal 2024 project titled "In Emergency, Break Glass."

It does not, however, focus on the Department of the Air Force's internal workforce or Space Force civilian employees; instead, it highlights shortfalls in the civilian sector.

RAND's report identifies shortfalls in three groups: engineers with bachelor's degrees or higher, skilled touch labor (technicians and assemblers), and advanced-degree specialists in electro-optical sensing and radiation hardening.

None of these are easy to replace or replicate in a short period, but one must consider that the 12- to 24-month target is one RAND has self-assessed in an industry that can take as much as a decade to replace satellites on average.

RAND notes that no such formal requirement exists, adding that the creation of the Space Force in 2019 could be read as a tacit acknowledgment of near-term dependence on space.

The people the report is most worried about, interestingly, are cleanroom technicians, welders, and solderers, not scientists, because of the labor shortages in play that already come to roughly nineteen jobs for one qualified individual.

This is further exacerbated by RAND essentially trying to set a somewhat aggressive standard that the Pentagon has yet to define formally. It also cited security clearance delays as a hiring bottleneck, with initial Secret investigations averaging more than 87 days in fiscal 2024 and Top Secret cases running close to six months.

Interestingly, RAND notes that in a conflict where the adversary can destroy systems faster than the United States can launch replacements, building exact replicas with known vulnerabilities may accomplish little, especially in a protracted war.

It also offers a startlingly small remedy for what it paints as a significant problem: RAND's recommendations cost an estimated $4 million a year for scholarships and internships, $1 million a year for outreach and recruitment, and a one-off $800,000 feasibility study on retraining workers from adjacent industries.

This is a minuscule amount compared to the roughly $32 billion the US Space Force is expected to have in play for Fiscal 2027, clocking in at just 0.016 percent of that figure, and 0.003 percent of the $185 billion earmarked for the Golden Dome project.

A war might not be required to exacerbate the issue, however; the White House's space transportation policy released in August targets more than 1,000 launches and reentries by 2030, up from roughly 200 a year today. A wartime surge without the requisite employees in training or in reserve could only further complicate those goals.

• Alien Deathstorm head of design Ben Fisher says the game's engineer protagonist "felt like a natural fit" in the game's environment and narrative
• He says Rebellion wanted to match an interesting character with "interesting mechanics" and landed on an engineer
• Producer Martin Willingham says "the motive was making the player feel immersed in the experience and exposed"

Rebellion has revealed the influences behind Alien Deathstorm and explained why it chose an engineer protagonist for its first-person survival horror game.

In an interview with TechRadar Gaming at Gamescom, producer Martin Willingham and head of design Ben Fisher shared that the team was, of course, inspired by the iconic Alien movies, but also looked at the broader influence of "90s cassette futurism" to create "that gritty, grounded sci-fi atmosphere."

"A lot of the tone came from our own 2008 comic series as well," Fisher said. "That kind of broad, pulpy, accessible sci-fi storytelling."

When discussing the intention behind having Deathstorm's protagonist be a protagonist instead of a typical soldier, Willingham explained that the studio wanted to find an interesting character that matched "interesting mechanics," and an engineer "felt like a natural fit."

The perspective also enabled Rebellion to "explore the elements of the sandbox gameplay," and an engineer working to diagnose the problems within a dangerous environment surrounded by monsters worked well.

"So part of the motive was making the player feel immersed in the experience and exposed," Fisher said, "because if you're an engineer trying to survive in that situation, you're less trained than a space marine or something like that [or] some kind of future soldier.

"You're going to feel more at threat. So it amplifies that feeling of being isolated, doing the best in a difficult situation."

While that's one aspect, the developer also said that the other is the added "moments of anticipation" that come with the engineer role, like the build-up you get in a horror experience while trying to complete your objective.

"You've got to slow down and observe the space, so you're not just rushing to the next objective," he added. "You're actually getting immersed in the experience, and that means that when you start hearing the roar of a distant alien, that split of attention builds up a whole load of anticipation and a load of fear."

Alien Deathstorm launches in Spring 2027 for PlayStation 5, Xbox Series X, Series S, and PC.

The GMKtec EVO-X3 is a mini PC with a difference: firstly, it stands upright; the size is larger than most, the specs are impressive, and the price- well, that prices it out of most people's reach; however, this is a machine with an AMD Halo Strix processor.

The EVO-X3 features the AMD Ryzen AI Max+ 395 processor; that’s the same as the EVO-X2 with 16 Zen 5 CPU cores, Radeon 8060S graphics, and a 50-TOPS XDNA 2 NPU. In this review, I’m looking at the machine with 128GB of LPDDR5X unified memory, enough to load and run local AI models; this ability is well beyond the capabilities of most other Mini PCs and even desktop PCs.

A quick look at the hardware makes it clear this is different, and for a machine with its specs, it’s still relatively compact, thanks to its slim, tower-like structure. It’s essentially been designed to stand upright to ensure plenty of airflow.

During testing, I ran local chat, reasoning, image recognition, image generation, transcription, and coding models on the EVO-X3. Smaller language models started responding almost immediately, while larger models provided more detailed reasoning and better-quality copy, essentially mirroring what you’d expect from cloud models.

You could also run these tasks with Wi-Fi disabled, so completely local and offline, with no need for credits or additional spend. This also meant all prompts, photographs, audio recordings, and project files stayed on your computer rather than being sent to a cloud service where you don't know what happens to them.

While the main focus here is the AI potential, I also ran the conventional benchmarks, and these are impressive. If you need a cutting-edge gaming PC and want to underplay the machine's potential, this is a great compact option, especially with the OcuLink connector.

However, refocusing on AI and unlocking the EVO-X3’s AI performance proved that some knowledge and experimentation are still needed; this isn’t quite an out-of-the-box, ready-to-run solution, although GMKtec has taken huge strides in that direction.

Local AI through the GMKtec Claw software environment was impressive. During testing, the Herdsman model manager recognised the NPU but reported that the Ryzen AI Max+ 395 was unsupported by its current NPU runtime manifests, which I found odd since the software was pre-installed.

Installing Lemonade and its Ryzen AI runtime solved the problem. Compatible models then pushed NPU utilisation close to 100%, proving that the hardware was operating properly.

This leaves the EVO-X3 in an awkward position. It’s an exceptionally powerful local AI workstation, but it is not yet the polished, out-of-the-box AI machine that the marketing suggests. It will do everything advertised, but for you to get there, you must understand models, runtimes, servers, agents and external applications, or at least be prepared to learn.

Whether or not this deserves a spot on our guide to the best mini PC is, ultimately pretty subjective. At $3,799.99 for the current 128GB/2 TB configuration, this is a relatively cheap machine for experienced developers, and technically minded creative professionals will find it an extraordinarily useful tool if they’re willing to put in the time. If you’re less experienced, then hold out; a more streamlined and cheaper option will inevitably be available soon.

• How much does it cost? From £3,029 / $3,799
• When is it out? Available now
• Where can you get it? Directly from GMKtec or Amazon

The GMKtec EVO-X3 is available in the US direct from GMKtec and on Amazon.com, with the base 128GB/2TB configuration priced at $3,800. However, you can save an extra $100 at GMKtec or 5% on Amazon with code GMKX3DP3.

In the UK, the EVO-X3 is priced at £3,030 from GMKtec. At the time of review, it's not currently on Amazon, but I'd expect it to arrive shortly.

The EVO-X3 comes with either a 2TB or 4TB SSD, with both versions featuring 128GB of onboard LPDDR5X-8000 memory.

• Value: 4 / 5

CPU: AMD Ryzen AI Max+ 395, 16 cores/32 threads, up to 5.1GHz
GPU: AMD Radeon 8060S, 40 compute units
NPU: AMD XDNA 2, 50 TOPS; up to 126 TOPS total AI performance
RAM: 128GB LPDDR5X-8000, non-upgradeable
Storage: 2TB or 4TB PCIe 4.0 M.2 2280 SSD, two slots, up to 16TB total
Front Ports: USB 3.2 Gen 1 5Gbps, USB 2.0, 3.5mm audio
Rear Ports: HDMI 2.1, USB4 40Gbps, USB 3.2 Gen 2 10Gbps, OCuLink PCIe 4.0 x4, 2.5GbE RJ-45
VESA Mount: Not specified
Dimensions: 13.90 by 7.32 by 1.61in (353 x 186 x 41mm)
Weight: 5.07lb (2.3kg)
Power: 230W external power adapter, 19.5V/11.8A
Operating System: Windows 11 Pro, with Ubuntu and Linux support

The EVO-X3 design is far removed from the usual square mini-PC, and instead goes for a tall, narrow metal enclosure with vents on either side that help draw cool air through the system to help maintain maximum performance.

While it’s tall, the footprint because of this is relatively small, meaning it can sit neatly next to your monitor. Standing at 353 x 186 x 41mm without its stand, it looks more like an oversized external graphics card than a conventional desktop computer; the small stand that’s included is held in place with two screws and adds stability to the vertical stance.

While the unit is slim, it does have a good weight at 2.3kg, so while you could transport this very easily, this is one designed to stay in situ.

The full-metal CNC-machined casing looks great once you get used to the slightly different design, and with the grey finish with green detailing, the aesthetics are all well balanced. Compared with the small boxy Mini PCs, this definitely looks and feels like a professional workstation, as the pricing suggests.

Taking a closer look at the cooling and the slim tower design enables air to be pulled in through the single vent plate on the top/side of the machine and output through the two vent panels on the base/side. Checking inside the casing, there are three heat pipes and multiple cooling fans that all help to keep the powerful components cool.

Through the test, I was impressed by just how well and how quietly these fans were and left on auto; while you can hear them, they’re never overbearing. Through the interface, you also have the ability to adjust the mode, from Silent 45W to Balanced 85W and Performance 140W.

One of my biggest surprises when it comes to the design is the port allocation, or lack of it. The front provides one USB-A 3.2 Gen 1 port, one USB-A 2.0 port and a 3.5mm audio connection. Considering the EVO-X3’s intended creative and professional audience, another high-speed USB connection or front-mounted USB-C port would have been handy.

The rear is slightly better equipped, with HDMI 2.1, 40Gbps USB4, 10Gbps USB-A, 2.5Gb Ethernet and an OCuLink connection. I would have preferred a little more, as connecting to a hub of some type is essential if you have further accessories.

The most interesting addition here is the OCuLink that enables you to add an external desktop graphics card via a more direct PCIe connection than a typical Thunderbolt or USB4 enclosure. Hot plugging isn't supported, so you must shut down the computer before attaching or removing an external GPU.

Wireless connectivity options are all pretty standard with Wi-Fi 7 and Bluetooth 5.4; then there’s also the wired 2.5Gb Ethernet. This is slightly lacking, and I would have expected a 5GbE or 10GbE at the least.

However, the lack of ports aside, the big story here is AMD’s Ryzen AI Max+ 395 that combines a 16-core, 32-thread Zen 5 CPU with Radeon 8060S integrated graphics. The GPU includes 40 RDNA 3.5 compute units, while the XDNA 2 NPU contributes up to 50 TOPS of dedicated AI processing.

Other machines with this CPU, such as the EVO-X2, they feature only 64GB of RAM and no Oculink connection, both of which, as I discovered, really are essential if you want to make the most of running local AI.

• Design: 4.5 / 5

At the heart of the EVO-X3 is AMD’s Ryzen AI Max+ 395, part of the AI power family Strix Halo. This chipset offers a combined 16-core, 32-thread Zen 5 CPU alongside Radeon 8060S graphics and an XDNA 2 NPU all within a single processor package, which is why it can fit into this small form Mini PC.

The CPU can be boosted to 5.1GHz and is supported by 16MB of L2 cache and 64MB of L3 cache. Alongside it, graphics processing is supplied by the Radeon 8060S, which provides 40 RDNA 3.5 compute units operating at up to 2,900MHz.

What is impressive here is that the performance of this integrated GPU has, in tests, beaten many dedicated GPUs. As GPU power is central to creative work, gaming and local generative AI, this power will be a huge benefit to the majority of Creative Apps that I use.

Alongside the CPU and GPU is the XDNA 2 NPU that provides up to 50 TOPS of dedicated AI processing, which brings the combined processor performance to up to 126 TOPS. The NPU essentially handles supported AI workloads better than running everything through the CPU or GPU. Again, surprisingly, some of the AI Apps used in the test fell back to the CPU and GPU rather than making use of the NPU and the speeds with and without made a big difference.

If you’re looking through this and thinking this all sounds very familiar, then it is essentially a very similar configuration to the EVO-X2, which is again equipped with the the Ryzen AI Max+ 395, Radeon 8060S graphics and a 50-TOPS NPU. The EVO-X3 adds the new chassis design to maximise the new cooling system, OCuLink connection, 128GB of standard memory capacity and a pre-installed AI environment to make getting started easier.

My review sample of the EVO-X3 features 128GB of onboard LPDDR5X memory running at 8,000MT/s. This memory is soldered and cannot be upgraded, but up to 96GB can be allocated as graphics memory. That large shared-memory pool is useful for local AI because it enables the integrated Radeon GPU to load models that would exceed the dedicated VRAM capacity of most graphics cards.

Storage options are either 2TB, as in my review sample, or 4TB M.2 PCIe 4.0 NVMe SSD. It includes two internal M.2 2280 PCIe 4.0 x4 slots, which enable you to install up to two 8TB drives, with a maximum internal capacity of 16TB. Having downloaded and installed several models, tools and applications through the test, I would definitely install a minimum of an additional 4TB.

One of the big upgrades over the EVO-X2 is the rear OCuLink connection. This provides a PCIe 4.0 x4 interface for an external desktop graphics card; you do, of course, need an eGPU enclosure as well, and then this gives you the ability to add an Nvidia or AMD GPU for rendering, gaming, as well as those AI workloads. OCuLink does not support hot plugging, so if you do use one, the machine needs to be powered down before the port is installed. I don’t at present have an OcuLink enclosure, so I haven’t tested the boost that this will give to performance.

On the back there’s also the USB4 connection, but only one, and it enables data transfers at up to 40Gbps, DisplayPort video output at up to 4K 60Hz and USB Power Delivery input at up to 100W. If I were to use this machine long-term, then I would connect a hub to maximise connections, especially for creative use, as just one USB-C is limiting.

Display output is limited to two screens, which can be connected through HDMI 2.1 and USB4. The HDMI connection supports higher-resolution displays, while the USB4 port provides DisplayPort output up to 4K at 60Hz.

Connectivity includes Wi-Fi 7, Bluetooth 5.4 and 2.5Gb Ethernet. The front has two 5Gbps USB-A 3.2 Gen 1 ports, a 3.5mm combined audio socket, and the power button. Around the back are 40Gbps USB4, 10Gbps USB-A 3.2 Gen 2, HDMI 2.1, OCuLink, 2.5Gb Ethernet and the main DC power input.

The absence of a front USB-C port or card reader is disappointing for such an expensive creative workstation. This is especially noticeable because the cheaper EVO-X2 includes both front USB4 and an SD card reader.

I was also pleased to see three performance profiles provide a choice between power consumption, heat and processing speed. Silent mode operates at 54W, Balanced mode increases this to 85W, and Performance mode allows sustained operation at up to 120W, with short peaks of up to 140W.

Three heat pipes, two fans, and intelligent fan control handle cooling by pulling air through the vent on the top/side and pushing it out through the bottom/side. This cooling remains relatively quiet throughout the testing and is designed to maintain the processor’s performance during heavy AI use, as well as video rendering and development workflows.

As you would expect, Windows 11 Pro is preinstalled, with support for Ubuntu and other Linux distributions. GMKtec also supplies its Claw local-AI environment, which is promoted as offering one-click model and agent deployment and was well implemented; there’s also Herdsman and installers for LM Studio and Ollama, giving you a good choice of Local AI management.

Claw is well integrated with EVO-X3 as it brings together local models, cloud services and AI agents without requiring you to assemble the entire software environment. The associated Herdsman model manager detected the NPU but reported that the Ryzen AI Max+ 395 was unsupported by its current runtime manifests. The underlying hardware worked once Lemonade and its Ryzen AI runtime were installed, but that workaround goes against GMKtec’s claim that local AI is ready to use without coding or setup.

The EVO-X3 has an exceptional feature set for local AI with the 128GB unified-memory capacity, integrated graphics, NPU, dual SSD slots and OCuLink connection enabling both flexibility and power. The issue is that most of this is already present in the substantially cheaper EVO-X2, making the X3’s £3,029.99 price seem a little steep.

• Features: 4.5 / 5

Crystal Disk Mark Read: 7045.28
Crystal Disk Mark Write: 6182.82
Geekbench CPU Multi: 9469
Geekbench CPU Single: 2619
Geekbench GPU: 25645
PC Mark Overall: 6855
Cinebench CPU Multi (Threads): N/a
Cinebench CPU Single (Threads): N/a
Fire Strike Overall: 7429
Fire Strike Graphics: 8551
Fire Strike Physics: 11509
Fire Strike Combined: 2953
Time Spy Overall: 3631
Time Spy Graphics: 3473
Time Spy CPU: 4897
Wild Life Overall: 24312
Steel Nomad Overall: N/a
Windows Experience Overall: 8.7

Getting started with the EVO-X3 is much like any other mini PC: plug in, screw the stand into the base and power on. You then have the Windows 11 Pro setup to run through, and as ever, GMKtec has covered the LAN port with a warning not to connect until the full setup process has been completed.

Once installed, I instantly noted the speed of the system; this was as fast and responsive as they come, and powering down and starting up from cold takes just 18 seconds from power being pushed to the home screen and the machine being ready to use.

As I started to load the applications for the benchmarks, the machine was fast and responsive, really as you’d expect for a machine at this price, and I was happy to see that this was at least in line, if not faster, than my workstation. Opening Premiere Pro, Lightroom and Photoshop all worked seamlessly, with the only real issue being the lack of ports for connection at the back.

When it came to the usual software and benchmarking tests, I ploughed through with almost all hitting record speeds and values when compared with any Mini PC or Laptop that I’ve reviewed in the past. An example is the Geekbench score, which aligned with many of the high-performance Mini PC’s and laptops, with a score of 2,956, but then the multi-core result of 19,401 far exceeded anything else.

Pushing the rendering benchmarks to give an idea of how it would perform for 3D content and video, the Cinebench R23 produced 33,295 points in the multi-core test and 2,034 points using a single core again; this is almost double anything else that I have seen.

Due to the positioning of the machine in the market, you would expect decent graphics performance, but due to an AI rather than games and creative perspective, and sure enough, the graphics performance was superb as I settled into some gaming. The Radeon 8060S hit scores of 101,583 in Geekbench GPU, while 3DMark Fire Strike returned an overall score of 25,721.

The EVO-X3 scored consistently high results through the 3DMark benchmarks, with most doubling the results that I have seen from other high-performance Mini PC. This performance not only translated to the test but also to gaming, as I settled back to play Cyber Punk 2097.

My review sample came with the 2TB NVMe SSD, and the transfer rates were slightly below what I expected but are consistent for a PCIe 4.0 slot, with read and write speeds of 4,912.96 MB/s and 4,648.62 MB/s.

Rounding out the standard tests, I checked out PCMark and wasn’t surprised to see a high score of 9,857, while the Windows Experience result was 9.5, again one of the highest I’ve recorded this year.
GMKtec EVO-X3 review: AI software

The EVO-X3 is sold as a ready-to-go machine with local AI models installed and one-click agent deployment through its Claw software environment. Claw uses Herdsman as the gateway to the AI models, so you start Herdsman and then open Claw, and the rest is quite straightforward for chat, image generation and countless other Agents that can be downloaded and installed.

Using this approach simplifies everything, and as I downloaded SuperPowers Agent, software workflow, Paper Research, Office Copilot, slidewright and a few others, I quickly built myself a decent set of tools for day-to-day AI use. The models you use can be downloaded in Herdsman and then used in GMKClaw. Depending on the complexity of the model you use and whether you switch Thinking on or off, the speed of the AI model you have selected varies.

What I like about this is that once you have models loaded, the application then downloads the additional tools that it needs to run. This is obviously online, with the AI processing done offline. Initially, the results were hit and miss, with the simple prompt, “A photo of a Labrador on a beach”, creating an abstract; once it had a few more prompts and details, the final results started to look better. The point here is that the more you use it, the better it gets.

While I also wanted to test the offline AI ability, what I also realised quickly was that I needed to run examples of everything I wanted to do; it needed to download and upgrade the models and tools that it had to hand.

Through the menu at the top of the screen, you can toggle between the different AI models, with Auto selected as default. This auto mode switches between the cloud and local to generate the images, but I noted that the NPU throughout stayed silent, registering 0% in Task Manager.

While the performance of the models was good both online, offline and hybrid, this wasn’t the potential that I was hoping for from this machine. I then went down another route and installed Lemonade and a series of Apps to give me greater control.

Installing Lemonade provided a way forward. After downloading its Ryzen AI NPU runtime and compatible models, Windows Task Manager showed the NPU operating at near-full potential, and a quick test in chat showed just how fast this machine's AI could be.

Lemonade provided what I had hoped for from Claw, with the clearest interface for testing local models; the GMKTecClaw integration is good but not yet mature enough to reach its full potential. Lemonade also displays time to first token, token-generation speed, memory use and processor activity, while making it relatively easy to switch between NPU, hybrid and GGUF models.

However, the setup process does require large model downloads and multiple tools and other additions to be downloaded. It’s also worth highlighting that downloading several models simultaneously resulted in HTTP 429 rate-limit messages from Hugging Face. Some completed downloads also encountered temporary file-locking problems, so for the complete set that I ran in this test, I downloaded over a few days, trying one task and then returning for more models for images, text, speech, and more.

Once Lemonade and its Ryzen AI runtime were installed, the EVO-X3’s NPU performance was impressive. I started out with the relatively lightweight Qwen-2.5-1.5B-Instruct-NPU model, which responded in around a second per response. Checking the stats, it recorded a time to first token of 0.61 seconds and generated 25.1 tokens per second. RAM usage reached 16.7GB, while CPU utilisation remained at only 4.8%.

This made it a good model for quick questions, summaries and relatively straightforward writing tasks.

Moving to Qwen2.5-7B-Instruct-NPU resulted in slower responses, as expected from a more complex model, but provided a useful step up in capability. The Qwen2.5-7B-Instruct-Hybrid model was again fast, and when asked to generate the opening of an imaginary book, it began responding in about a second, with Lemonade showing 523 input tokens, 393 output tokens, and a time to first token of 0.47 seconds.

I then loaded the larger gpt-oss-20b-NPU, more brain and more thinking, so everything took longer to consider when it came to the responses but produced noticeably better copy. It did, however, become slightly preoccupied with satisfying the requested word count and took quite some time rewriting the 100 words. That test processed 825 input tokens and generated 1,500 output tokens. Time to first token was 1.67 seconds, while RAM use reached 25.6GB and CPU remained at 5.9%.

These tests show why model selection matters. The smallest model delivered the fastest chatbot for general use, while larger models slowed speed for improved reasoning and writing quality. All processing was run locally, disconnected from the internet.

The next step was to plan a project, or, to be more accurate, all the things that I had to get done today. I wanted a plan and a spreadsheet that I could open on my phone and also print out if needed. I used DeepSeek-R1-Distill-Qwen-7B-NPU to organise the workday and plan an entire website project, covering content gathering, site structure, visual design, and implementation.

After asking several questions to clarify the project requirements, the model produced a detailed, methodical plan with useful software suggestions.

The website-planning conversation used 4,118 input tokens and generated 490 output tokens. Time to first token was 3.38 seconds, with RAM use reaching 19GB and CPU utilisation at 5.2%.

Generating the plan was straightforward, but turning it into a downloadable CSV proved difficult in the basic chat interface. While the model could create correctly formatted CSV text, it couldn’t save it as a downloadable file. Copying and pasting the information worked, but this highlighted the difference between a chat model and an agent with access to files and tools, which is where I started to discover additional tools needed to be used; with coding later, Visual Studio Code and the Continue extension were invaluable.

Essentially, the chat model can provide instructions and generate content, while an agent needs permission to create files, organise folders, run applications and modify a project directly, and everything needs connecting, which can take time initially, but after it’s done, the potential is exactly what you expect.

Image recognition is something that I’m really interested in, and training a model to help organise thousands of files literally. To start this off, I loaded a photograph into Gemma-3-4b-it-GGUF and asked it to describe the person shown, which was me.

The model correctly recognised that the subject was male and produced a reasonably accurate description of my appearance and clothing. It processed 267 input tokens and generated 147 output tokens. Time to first token was considerably slower at 22.65 seconds. The process used 17.2GB of RAM, while CPU utilisation remained at approximately 4.6%.

When I ran the UL Procyon AI Computer Vision benchmark, it returned a score of 1,171, which highlights this machine's potential on this front.

Through this part of the test, I looked at using local AI to organise an archive containing thousands of photographs. Gemma suggested several workable approaches, requiring Python scripts and external utilities rather than simply selecting a folder and pressing a button. Again, if you have more complex tasks, then these are possible but do require you to set up the resources required.

As part of this test, I installed ExifTool and ImageMagick, which provided the underlying ability to read metadata, organise files, and produce contact sheets, but connecting these tools to an agent required more configuration.

The EVO-X3 was powerful enough for the job. The difficulty was connecting the model to the required tools and granting it the ability to examine, classify and move files safely. This requires your knowledge and skill and is one of the features that is possible but is not ready out of the box.

Video organisation presents the same problem. Switching models provides different abilities, but no single supplied application will automatically understand, catalogue, and reorganise a large photo or video archive.

Local image generation produced surprisingly good results. I wrote several detailed prompts and used them to create original images.

The generated pictures followed the descriptions closely and, while each took approximately two minutes and 30 seconds to render, the resulting quality was more than acceptable for experimentation, visualisation and concept development.

During one test, Lemonade displayed 17 input tokens, 509 output tokens and a 0.25-second initial response. RAM usage reached 18.1GB, while CPU utilisation was 6.9%.

The token figures are less useful than the total rendering time because image generation differs considerably from language-model inference. However, the test showed that the EVO-X3 can produce useful generative imagery without a discrete GPU or cloud-based service.

I couldn't complete a comparable video-generation test. This would have required another specialist model and a separate workflow, such as ComfyUI, which would have extended the setup beyond the applications already installed.

Another of my major uses recently of AI agents is transcription. The UGREEN IDX6011 features a simple, ready-to-go app, and this machine looks to offer the same. To get started, I loaded a 46-minute podcast about 3D printing into Whisper-Large-v3-Turbo. The interface let me select the audio file but didn't provide a prompt field, preferring to just get on with it. As processing progressed, Windows Task Manager showed repeated spikes in NPU activity, showing that the dedicated accelerator was being used.

The processed transcription appeared after about five minutes and was really accurate. The speed and accuracy of this transcription are outstanding, and I’ll look in the near future to see if time codes can easily be added, which will help with editing. Again, all local and not cloud intervention.

AI code creation is one of the big areas of AI, and having used this machine for a month, I can see why. It’s not quite as straightforward as design me this site, and then it will. You need to provide it with the guidance, but as long as you do that, you can essentially bash together a site relatively quickly.

I still couldn’t get it to perform quite as well as some of my other workflows from Figma to GitHub and Cloudflare Pages, but still the workflow was quick and pretty solid when using the Qwen2.5-Coder-7B-Instruct model and connecting it through Visual Studio Code using the Continue extension.

As I created a site, the NPU utilisation reached 99% as the model generated the website structure and files. Further prompts could then add content, adjust the design, and fix issues. There was quite a bit of initial setup, and the end results needed an aesthetic polish, but for an hour's work I had a fully functional static site that could be updated quickly.

This test highlighted how a practical local-AI workflow, along with extensions and applications, can be used together to get the results that you want. Lemonade or Herdsman loads and serves the model, such as Qwen, providing the coding capability; Visual Studio Code then supplies the workspace and Continue enables the model to interact with the project.

• Performance: 4.5 / 5

The GMKtec EVO-X3 is a mini PC workstation and one that has been fine tuned to run local AI properly. The Ryzen AI Max+ 395 processor, Radeon 8060S graphics and 128GB of unified memory are provide the speed that's required and while more advanced models do push the system and slow things down,at least you get to use it without incurring cost.

Side by side the localised models aren't as powerful as the cloud based, but still what you're able to do is impressive and if you invest the time to develop your own AI workflows the end result will be AI agents in this local machine that can help with your workflows.

The bundled software is impressive, although there are some glitches at present that do slow the processing, however, a switch to another AI management option quickly unlocks speed and greater potential.

Out of the box this is one of the most impressive AI machine that really can run decent size local models. If you're willing to put in the time to discover how everything interlinks then this is a formidable machine with huge potential.

Value

Exceptional hardware, but extremely expensive 

4

Design

Solid construction, compact footprint and superb quiet cooling

4.5

Features

128GB unified-memory capacity, USB4 and OCuLink 

4.5

Performance

Outstanding CPU, graphics and local-AI capability once configured

4.5

Overall

Superb workstation hardware held back by setup complexity

4

You need to run substantial AI models locally.

The combination of Radeon graphics, a 50-TOPS NPU, and a large unified-memory pool enables models to run without relying on cloud processing.

You work with sensitive material.

You can process photographs, audio, documents, and source code locally without uploading them to a third-party service.

You expect an appliance-like AI experience.

The bundled model manager did not initially support the NPU, and accessing the hardware required Lemonade and separately downloaded models.

You only need a conventional mini PC

This level of processing, graphics, memory and AI performance is unnecessary for normal office work and occasional creative use.

For more professional-grade desktops, we've tested the best business computers.