One-bit large language models (LLMs) have emerged as a promising approach to making generative AI more accessible and affordable. By representing model weights with a very limited number of bits, 1-bit LLMs dramatically reduce the memory and computational resources required to run them.

Microsoft Research has been pushing the boundaries of 1-bit LLMs with its BitNet architecture. In a new paper, the researchers introduce BitNet a4.8, a new technique that further improves the efficiency of 1-bit LLMs without sacrificing their performance.

The rise of 1-bit LLMs

Traditional LLMs use 16-bit floating-point numbers (FP16) to represent their parameters. This requires a lot of memory and compute resources, which limits the accessibility and deployment options for LLMs. One-bit LLMs address this challenge by drastically reducing the precision of model weights while matching the performance of full-precision models.

Previous BitNet models used 1.58-bit values (-1, 0, 1) to represent model weights and 8-bit values for activations. This approach significantly reduced memory and I/O costs, but the computational cost of matrix multiplications remained a bottleneck, and optimizing neural networks with extremely low-bit parameters is challenging. 

Two techniques help to address this problem. Sparsification reduces the number of computations by pruning activations with smaller magnitudes. This is particularly useful in LLMs because activation values tend to have a long-tailed distribution, with a few very large values and many small ones.  

Quantization, on the other hand, uses a smaller number of bits to represent activations, reducing the computational and memory cost of processing them. However, simply lowering the precision of activations can lead to significant quantization errors and performance degradation.

Furthermore, combining sparsification and quantization is challenging, and presents special problems when training 1-bit LLMs. 

“Both quantization and sparsification introduce non-differentiable operations, making gradient computation during training particularly challenging,” Furu Wei, Partner Research Manager at Microsoft Research, told VentureBeat.

Gradient computation is essential for calculating errors and updating parameters when training neural networks. The researchers also had to ensure that their techniques could be implemented efficiently on existing hardware while maintaining the benefits of both sparsification and quantization.

BitNet a4.8

BitNet a4.8 addresses the challenges of optimizing 1-bit LLMs through what the researchers describe as “hybrid quantization and sparsification.” They achieved this by designing an architecture that selectively applies quantization or sparsification to different components of the model based on the specific distribution pattern of activations. The architecture uses 4-bit activations for inputs to attention and feed-forward network (FFN) layers. It uses sparsification with 8 bits for intermediate states, keeping only the top 55% of the parameters. The architecture is also optimized to take advantage of existing hardware.

“With BitNet b1.58, the inference bottleneck of 1-bit LLMs switches from memory/IO to computation, which is constrained by the activation bits (i.e., 8-bit in BitNet b1.58),” Wei said. “In BitNet a4.8, we push the activation bits to 4-bit so that we can leverage 4-bit kernels (e.g., INT4/FP4) to bring 2x speed up for LLM inference on the GPU devices. The combination of 1-bit model weights from BitNet b1.58 and 4-bit activations from BitNet a4.8 effectively addresses both memory/IO and computational constraints in LLM inference.”

BitNet a4.8 also uses 3-bit values to represent the key (K) and value (V) states in the attention mechanism. The KV cache is a crucial component of transformer models. It stores the representations of previous tokens in the sequence. By lowering the precision of KV cache values, BitNet a4.8 further reduces memory requirements, especially when dealing with long sequences. 

The promise of BitNet a4.8

Experimental results show that BitNet a4.8 delivers performance comparable to its predecessor BitNet b1.58 while using less compute and memory.

Compared to full-precision Llama models, BitNet a4.8 reduces memory usage by a factor of 10 and achieves 4x speedup. Compared to BitNet b1.58, it achieves a 2x speedup through 4-bit activation kernels. But the design can deliver much more.

“The estimated computation improvement is based on the existing hardware (GPU),” Wei said. “With hardware specifically optimized for 1-bit LLMs, the computation improvements can be significantly enhanced. BitNet introduces a new computation paradigm that minimizes the need for matrix multiplication, a primary focus in current hardware design optimization.”

The efficiency of BitNet a4.8 makes it particularly suited for deploying LLMs at the edge and on resource-constrained devices. This can have important implications for privacy and security. By enabling on-device LLMs, users can benefit from the power of these models without needing to send their data to the cloud.

Wei and his team are continuing their work on 1-bit LLMs.

“We continue to advance our research and vision for the era of 1-bit LLMs,” Wei said. “While our current focus is on model architecture and software support (i.e., bitnet.cpp), we aim to explore the co-design and co-evolution of model architecture and hardware to fully unlock the potential of 1-bit LLMs.”

Starfish Space has closed a new tranche of funding led by a major defense tech investor as it looks to launch three full-size satellite servicing and inspection spacecraft in 2026. 

The Washington-based startup’s Otter spacecraft is designed for two primary missions: extending the operational life of expensive satellites in geostationary orbit (GEO) and disposing of defunct satellites in low Earth orbit (LEO). It’s a series of capabilities that have never been available for satellite operators, who launch their satellites with the expectation that they’ll only have a limited span of useful life. 

The aim, as Starfish CEO and co-founder Austin Link put it in a recent interview, is to “make it affordable enough that the benefits of having your satellite serviced outweigh the costs.”

The $29 million round was led by Shield Capital, a venture firm focused on funding technologies that will affect U.S. national security. It has participated in just a handful of other deals in the space industry. The round also includes participation from new investors Point72 Ventures, Booz Allen Ventures, Aero X Ventures, Trousdale Ventures, TRAC VC, and existing investors Munich Re Ventures, Toyota Ventures, NFX, and Industrious Ventures.

“You start a company because you want to build satellites, not because you want to fundraise,” Link told TechCrunch. Link founded Starfish in 2019 with Trevor Bennett after the pair worked as flight sciences engineers at Blue Origin. They raised $7 million in 2021 and $14 million two years later. Starfish launched its first demonstration mission, a sub-scale spacecraft fittingly called Otter Pup, last summer. 

Although that mission did not quite go according to plan, Starfish has racked up several wins since then, including three separate contracts for full-size Otter spacecraft. That includes a $37.5 million deal with the U.S. Space Force for a first-of-its-kind docking and maneuvering mission with a defense satellite in GEO and a contract with major satellite communications company Intelsat for life extension services. The third contract, a $15 million NASA mission to inspect multiple defunct satellites in LEO, was announced while Starfish was in the middle of fundraising, Link said. 

Starfish purposefully set out to find investors that had experience helping their portfolio companies navigate selling to the government, Link said. “The government is a customer that it sometimes can be harder to scale with, so having investors that understood the process a little better … we thought they’d be good additions to our cap table.”

Link added that the company is seeing a “fairly even split” in demand between government and commercial customers. 

Satellite servicing, life extension, and satellite disposal are “exciting first steps,” Link said, but they’re stepping stones on the way to developing a broader suite of capabilities for even more ambitious missions on orbit. 

“Along the way, we end up with this set of autonomy and robotics technologies and capabilities and datasets that allow us to go eventually do broadly a set of complex robotic or servicing or ISAM-type missions in space that maybe stretch a little beyond what we do with the Otter,” he said. “I think a lot of those are a long ways off, and not necessarily where our focus is right now … but some of the effort that goes into the Otter today and is funded through this funding round, and some of the growth there leads to a longer term where Starfish Space can have a broad impact on the way that humans go out into the universe.”  

Apple today announced some minor updates to Logic Pro for both the Mac and the iPad, including the ability to search for plug-ins and sources and the addition of more analog-simulating sounds.

In Logic Pro for Mac 11.1 and Logic Pro for iPad 2.1, you can now reorder channel strips and plug-ins in the mixer and plug-in windows to make it easier to organize the layout of an audio mix.

As for the new sounds, Apple added a library of analog synthesizer samples called Modular Melodies, akin to the Modular Rhythms pack already found in Logic.

A more exciting sonic addition is the new Quantec Room Simulator (QRS) plug-in, which emulates the vintage digital reverb hardware of the same name, found in professional recording studios all over the world. Apple has acquired the technology for the classic QRS model and the later YardStick models to integrate into this software.

Specific to Logic Pro for Mac, you are now able to share a song to the Mac’s Voice Memos app — which may be a great feature for when Voice Memos gets that multitrack option on the iPhone in iOS 18.2

Added to the iPad version of Logic Pro is the ability to add your own local third party sample folders to the browser window, to make it easier to bring external audio files into tracks and sampler plug-ins.

These upgrades are small for current Logic users, but they do overall make the digital audio workplace easier to use and adds to the plethora of useful tools for no additional cost. Users will have access to upgrade to Logic Pro for Mac 11.1 and Logic Pro for iPad 2.1 today.

Having a very, very long Wordle streak is a blessing and a curse. On the one hand, it grants me immense bragging rights over those mere mortals with their streaks in the hundreds. On the other, it turns every game into a must-win ordeal. After all, what would I be without my Wordle streak? A so-called ‘expert’ with no credentials, that’s what. I’d be laughed out of town.

I joke, of course, but having gone 1,046 games without a loss I would rather not give up my streak all the same. I nearly had to today, though, because game #1,244 (Thursday, 14 November) nearly sent me back to ground zero.

Getty Images

A price to pay

‘Very high value damages’

If you’re in the market for a new Mac (or Apple display), there’s a lot of choice ahead of you. Maybe you’re interested in a lightweight MacBook Air from the selection of the best MacBooks — or maybe one of the desktop Macs. Either way, there’s a wide variety of Apple products on offer, including some external desktop monitors.

Below you’ll find the latest information on each model, including if it’s a good time to buy and when the next one up is coming.

MacBook Pro

The MacBook Pro represents the high-end of Apple’s MacBook lineup, and it was updated it in October 2024. This revamp brought the M4 chip series to the flagship laptop, including the M4, the M4 Pro, and the M4 Max.

However, it was much more than just a chip update, as Apple increased the memory capacity at every level (including double the starting RAM of the previous MacBook Pro). There’s a black color option for the entry-level M4 MacBook Pro, a brighter display with a nano-texture option, an improved webcam, more ports with Thunderbolt 4 connectivity, plus longer battery life. And let’s not forget the mini-LED display that can range up to 1,600 nits of peak HDR brightness and 1,000 of SDR brightness.

Given the recent update, we wouldn’t expect a new MacBook Pro until at least fall 2025, which would fit in with Apple’s rumored plans to update this laptop on a regular, annual basis. This new model is likely to be a modest upgrade, with a much larger refresh rumored for 2026. This could get an OLED display and a thinner chassis, while Apple also is expected to bring out a foldable MacBook at some point in the coming years.

MacBook Air

Apple’s slimmest laptop received its most recent changes in March 2024 when Apple launched the M3 MacBook Air. The M3 chip brings hardware-accelerated ray tracing and better performance compared to the previous MacBook Air, as well as better microphones and Wi-Fi 6E connectivity. That’s wrapped up in Apple’s thinnest and lightest laptop chassis, along with a 500-nit display and comfortable keyboard.

The MacBook Air is the next Mac in line for an update, with a new version expected to launch in spring 2025. While we can be certain that this will come with the M4 chip, there’s not a lot else we know for sure about it. There are murmurings that Apple is working on an OLED MacBook Air, but this isn’t due to launch for a few years.

Mac mini

In October 2024, Apple gave the Mac mini its most significant overhaul in 14 years. The device’s case was completely redesigned to take on the appearance of a downsized Mac Studio, with a footprint about half the size of the previous model. Apple skipped the M3 chip entirely and upgraded the Mac mini from the M2 to the M4 and M4 Pro chips, which has resulted in an impressive performance leap. Elsewhere, the memory capacity has been increased at every tier, there’s support for up to three external displays, there’s a new port arrangement that features Thunderbolt 5 on the M4 Pro model, and you get two USB-C ports on the front of the computer (the USB-A ports have been removed).

Apple won’t bring out another Mac mini until fall 2025 at the earliest. There’s not much in the way of rumors for future Mac mini models, but we’ll update this page as soon as we hear more.

iMac

In October 2024, the iMac became the first Apple computer to get the M4 chip (following the iPad Pro in May 2024). Aside from the speedy new chip, you can now buy it in a new range of colors, with increased memory bandwidth, more starting memory, an improved webcam, as well as Apple’s nano-texture glass option that cuts down on reflections and glare (this was previously only available on the Pro Display XDR and the iPad Pro).

As with several other M4 Macs, the iMac’s next update is likely to arrive around fall 2025. There is talk that Apple is working on a larger iMac, perhaps with a 30-inch or 32-inch display and mini-LED tech, and this might debut alongside the next iMac in 2025. That said, this project has reportedly been delayed many times, so don’t bet the house on it arriving any time soon.

Mac Studio

The latest update to the Mac Studio arrived in June 2023. Today, it comes with M2 Max and M2 Ultra chips, huge amounts of memory, wide support for its external displays, as well as compatibility with high-impedance headphones, Wi-Fi 6E, and Bluetooth 5.3. It’s a highly capable computer inside a very compact chassis, making it ideal for high-end workloads even if you don’t have a lot of desk space.

Apple has skipped the M3 generation in the Mac Studio, so its next update should bring the M4 Max and M4 Ultra chips alongside a significant increase in performance. Reports have suggested that the M4 Ultra will have 32 CPU cores and 80 GPU cores, double the core counts you can find in the MacBook Pro’s M4 Max chip.

Mac Pro

Like the Mac Studio, the Mac Pro received its most recent update in June 2023. This was a sizable change for Apple’s most powerful Mac, as it was the first time it came outfitted with Apple silicon chips (in this case, the M2 Ultra). Buying this model today gets you the most modular (and expensive) Mac on the market, including support for PCIe expansion, huge amounts of memory, massive support for external displays, and high-end CPU and GPU performance.

Looking ahead, the M4 Ultra is a certainty, and this is expected to arrive in mid-2025, according to Bloomberg reporter Mark Gurman. There are rumors that this might result in the Mac Pro getting a maximum memory capacity of 512GB, up from the current 192GB.

There’s also the question of the Mac Pro’s design. The current chassis was designed for the 2019 model, which featured Intel chips that ran hot, as well as expandable, discrete graphics cards. That’s no longer the case, and it’s arguable that the Mac Pro doesn’t need to be as large or as focused on cooling as it currently is now that it uses Apple silicon chips. There aren’t many rumors that Apple will soon redesign the Mac Pro’s chassis, so don’t expect that to come next year, but don’t be surprised if it’s somewhere on the horizon.

Apple Studio Display

The 27-inch Studio Display is Apple’s more accessible monitor (at least compared to the Pro Display XDR, anyway). It comes with a 5K resolution and 600 nits of brightness, an internal A13 Bionic chip that handles Center Stage and Spatial Audio, a nano-texture glass option that cuts down on reflections and glare, one Thunderbolt 3 slot and three USB-C ports, plus two different stand options.

There’s a bit of confusion surrounding the date we can expect a new version of the Studio Display. Apple analyst Ming-Chi Kuo believes we could see it in 2025 or 2024 (the latter date being very unlikely now), whereas display industry expert Ross Young thinks Apple has suspended its plans for a Studio Display with a 120Hz ProMotion refresh rate. Aside from that, not much else is known about what we might see in the next Studio Display.

Pro Display XDR

We first saw the Pro Display XDR in 2019, when it launched alongside the then-redesigned Mac Pro. It’s Apple’s most capable monitor, and it comes with a 32-inch 6K screen, up to 1,600 nits of peak brightness (for HDR content only), 576 local dimming zones, a nano-texture glass option, one Thunderbolt 3 and three USB-C ports, and two stand options.

Out of all the products in this article, the Pro Display XDR has gone the longest without an update, as it hasn’t been changed since it launched in 2019. Seeing as it was released in tandem with the Mac Pro, we’d expect the same to happen this time around, and with the Mac Pro slated for a 2025 refresh, the Pro Display XDR might be updated at the same time. Rumors have suggested that it will get an onboard chip just like the Studio Display, which should bring a few more software features to the table. Other than that, not much else is known at this point.

Advertisement