Premium Content Waitlist Banner

Digital Product Studio

1 Minute News new

Nunchaku v0.1.4 Unleashed SVDQuant To 3x Speed Up 12B FLUX On RTX 4090

The world of AI image generation is rapidly evolving, with diffusion models leading the charge. However, the computational demands of these powerful models can be a significant bottleneck. That’s where Nunchaku v0.1.4 comes in, offering a groundbreaking solution for accelerating diffusion model inference. This release introduces SVDQuant, a novel quantization technique that drastically reduces memory usage and boosts speed, all while maintaining impressive image quality.

Table of contents

What is Nunchaku?

Nunchaku is an inference engine designed to optimize and accelerate diffusion models. The latest version, v0.1.4, leverages a cutting-edge technique called SVDQuant to achieve remarkable performance gains. The name itself, Nunchaku, evokes the speed and efficiency of the martial arts weapon, reflecting the project’s focus on rapid and streamlined inference. You can find the inference engine at Github.

The Challenge: Scaling Diffusion Models

Diffusion models, like FLUX.1 and PixArt-∑, are capable of generating incredibly detailed and realistic images. As these models scale up in size (reaching billions of parameters), the computational requirements become immense. This leads to:

  • High memory usage.
  • Longer processing times.
  • Difficulty deploying on resource-constrained devices (like laptops).

Traditional approaches to model optimization, like quantization, have focused primarily on reducing the size of model weights. However, diffusion models are often computationally bound, meaning that simply reducing weight size doesn’t translate to significant speed improvements.

What is SVDQuant and How Does It Work?

SVDQuant (Singular Value Decomposition Quantization) is a post-training quantization technique that addresses the limitations of previous methods. It achieves this by:

  1. Quantizing both weights and activations: Unlike methods that focus solely on weights, SVDQuant aggressively quantizes both weights and activations to 4 bits. This is crucial for achieving real speedups in computationally bound models.
  2. Absorbing Outliers with a Low-Rank Branch: A major challenge with 4-bit quantization is the presence of outliers in the values of weights and activations. SVDQuant cleverly uses Singular Value Decomposition (SVD) to decompose the weight matrix into a low-rank component and a residual.
    • The low-rank component, run at 16-bit precision, captures the most significant information (including the outliers).
    • The residual component can then be safely quantized to 4 bits.
  3. Kernel Fusion with Nunchaku: The Nunchaku inference engine is co-designed with SVDQuant. It fuses the operations of the low-rank branch with the 4-bit quantization kernels, minimizing memory access and kernel call overhead. This is key for realizing practical speedups.

Nunchaku Performance and Quality: The Best of Both Worlds

The results of SVDQuant and Nunchaku are impressive:

  • 3.6x Memory Reduction: The model size of the 12B FLUX.1 model is reduced by 3.6x compared to the BF16 model.
  • Up to 10.1x Speedup: On a 16GB laptop 4090 GPU, Nunchaku achieves up to a 10.1x speedup by eliminating CPU offloading and outperforming the NF4 W4A16 baseline by 3x.
  • 3.5 Memory Usage reduction: Nunchaku reduces memory by 3.5 times compared to 16-bit.

Nunchaku Image Quality: Maintaining Visual Fidelity

Despite the aggressive 4-bit quantization, SVDQuant maintains excellent image quality:

  • Outperforms Baselines: The 4-bit models generated with SVDQuant outperform NF4 W4A16 baselines on FLUX.1 models, exhibiting better text alignment and closer resemblance to the 16-bit models.
  • Superior to W4A8: On PixArt-∑ and SDXL-Turbo, SVDQuant’s 4-bit results surpass the visual quality of other W4A4 and even W4A8 baselines.
  • LoRA Compatibility: SVDQuant seamlessly integrates with LoRA (Low-Rank Adaptation), allowing for style customization without requantization.

Installation and Usage

Getting started with Nunchaku is straightforward. The installation process varies slightly depending on your operating system and environment:

Wheels (for Linux and Windows WSL):

The easiest way to install is using pre-built wheels:

pip install torch==2.6 torchvision==0.21 torchaudio==2.6  # Example: Install PyTorch 2.6
pip install https://huggingface.co/mit-han-lab/nunchaku/resolve/main/nunchaku-0.1.4+torch2.6-cp311-cp311-linux_x86_64.whl # Example: For Python 3.11 and PyTorch 2.6

Make sure you have PyTorch >= 2.5, and replace with link from huggingface.

Build from Source:

For more control or specific configurations, you can build from source:

conda create -n nunchaku python=3.11
conda activate nunchaku
pip install torch torchvision torchaudio
pip install ninja wheel diffusers transformers accelerate sentencepiece protobuf huggingface_hub
pip install peft opencv-python gradio spaces GPUtil  # For gradio demos
#Optional NVFP4 on Blackwell
pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128

#Install make sure g++ >=11
conda install -c conda-forge gxx=11 gcc=11
git clone https://github.com/mit-han-lab/nunchaku.git
cd nunchaku
git submodule init
git submodule update
pip install -e . --no-build-isolation
  • Ensure your CUDA version is ≥ 12.2 on Linux and ≥ 12.6 on Windows.
  • Currently supports NVIDIA GPUs with architectures sm_86 (Ampere: RTX 3090, A6000), sm_89 (Ada: RTX 4090), and sm_80 (A100).

Conclusion: The Future of Efficient Diffusion Models With Nunchaku 

Nunchaku v0.1.4, powered by SVDQuant, represents a significant step forward in making large diffusion models more accessible and efficient. By enabling 4-bit quantization of both weights and activations, and through clever kernel fusion, Nunchaku unlocks substantial memory savings and impressive speedups without sacrificing image quality. This opens up new possibilities for deploying these powerful models on a wider range of devices and accelerating research in the field. The Nunchaku project demonstrates that high-performance, low-precision inference is achievable even for the most demanding AI models. The Nunchaku engine provides speed and optimization.

| Latest From Us

SUBSCRIBE TO OUR NEWSLETTER

Stay updated with the latest news and exclusive offers!


* indicates required
Picture of Faizan Ali Naqvi
Faizan Ali Naqvi

Research is my hobby and I love to learn new skills. I make sure that every piece of content that you read on this blog is easy to understand and fact checked!

Leave a Reply

Your email address will not be published. Required fields are marked *

Forget Towers: Verizon and AST SpaceMobile Are Launching Cellular Service From Space

Imagine a future where dead zones cease to exist, and geographical location no longer dictates connectivity access. This ambitious goal moves closer to reality following a monumental agreement between a major US carrier and a burgeoning space-based network provider.

Table of Contents

Verizon (VZ) has officially entered into a deal with AST SpaceMobile (ASTS) to begin providing cellular service directly from space starting next year.

This collaboration signals a significant step forward in extending high-quality mobile network coverage across the U.S., leveraging the unique capabilities of satellite technology.

Key Takeaways

  • Verizon and AST SpaceMobile signed a deal to launch cellular service from space, commencing next year.
  • The agreement expands coverage using Verizon’s 850 MHz low-band spectrum and AST SpaceMobile’s licensed spectrum.
  • AST SpaceMobile shares surged over 10% before the market opened Wednesday following the deal announcement.
  • The partnership arrived two days after Verizon named Dan Schulman, the former PayPal CEO, as its new Chief Executive Officer.

Verizon AST SpaceMobile Cellular Service Launches Next Year

Verizon formally signed an agreement with AST SpaceMobile (ASTS) to launch cellular service from space, with services scheduled to begin next year.

Infographic

This announcement, updated on Wednesday, October 8, 2025, confirmed a major step forward for space-based broadband technology. The deal expands upon a strategic partnership that the two companies originally announced in early 2024.

While the collaboration details are public, the financial terms of the agreement were not disclosed by either party. This partnership is crucial for Verizon as it seeks to extend the scope and reliability of its existing network coverage.

Integrating the expansive terrestrial network with innovative space-based technology represents a key strategic direction for the telecommunications giant.

Integrating 850 MHz Low-Band Spectrum for Ubiquitous Reach

A core component of the agreement involves leveraging Verizon’s licensed assets to maximize the reach of the new system. Specifically, the agreement will extend the scope of Verizon’s 850 MHz premium low-band spectrum into areas of the U.S.

that currently benefit less from terrestrial broadband technology, according to rcrwireless.

This low-band frequency is highly effective for wide-area coverage and penetration.

AST SpaceMobile’s network provides the necessary infrastructure for this extension, designed to operate across several spectrums, including its own licensed L-band and S-band.

Furthermore, the space-based cellular broadband network can handle up to 1,150 MHz of mobile network operator partners’ low- and mid-band spectrum worldwide, the company stated. This diverse spectrum utilization ensures robust, global connectivity.

Abel Avellan, founder, chairman, and CEO of AST SpaceMobile, emphasized the goal of this technical integration. He confirmed the move benefits areas that require the “ubiquitous reach of space-based broadband technology,” specifically enabled by integrating Verizon’s 850 MHz spectrum.

Market Reaction and Verizon’s CEO Transition

The announcement immediately generated a strong positive reaction in the market for AST SpaceMobile.

Shares of AST SpaceMobile, which operates the space-based cellular broadband network, soared more than 10% before the market opened Wednesday, reflecting investor confidence in the partnership as reported on seekingalpha.com.

This surge indicates the perceived value of collaborating with a major carrier like Verizon to accelerate the deployment of space technology.

The deal arrived just two days after Verizon announced a major shift in its executive leadership. The New York company named former PayPal CEO Dan Schulman to its top job, taking over the post from long-time Verizon CEO Hans Vestberg.

Schulman, who served as a Verizon board member since 2018 and acted as its lead independent director, became CEO immediately.

Vestberg will remain a Verizon board member until the 2026 annual meeting and will serve as a special adviser through October 4, 2026.

This high-profile corporate transition coincided closely with the launch of the strategic Verizon AST SpaceMobile cellular initiative, positioning the service expansion as a key priority under the new leadership structure.

Paving the Way for Ubiquitous Connectivity

The ultimate vision driving this partnership centers on achieving truly ubiquitous connectivity across all geographies. Srini Kalapala, Verizon’s senior vice president of technology and product development, highlighted the impact of linking the two infrastructures.

He stated that the integration of Verizon’s “expansive, reliable, robust terrestrial network with this innovative space-based technology” paves the way for a future where everything and everyone can be connected, regardless of geography.

Leveraging low-band spectrum for satellite service provides a critical advantage in covering vast, underserved territories. The design of SpaceMobile’s network facilitates service across various licensed bands, maximizing compatibility and reach.

This approach ensures customers can utilize the space-based broadband without interruption, enhancing service quality in remote or challenging areas.

Conclusion: The Future of Verizon AST SpaceMobile Cellular Service

The agreement between Verizon and AST SpaceMobile sets a clear timeline for the commercialization of cellular service from space, beginning next year.

By combining Verizon’s premium 850 MHz low-band spectrum with AST SpaceMobile’s specialized satellite capabilities, the partners aim to dramatically improve broadband reach across the U.S.

This initiative demonstrates a powerful commitment to eliminating connectivity gaps, fulfilling the stated goal of connecting people regardless of their physical location.

The soaring stock value for AST SpaceMobile following the announcement underscores the market’s enthusiasm for this technological fusion.

Furthermore, the simultaneous leadership transition to Dan Schulman suggests this strategic space-based expansion will feature prominently in Verizon’s near-term development goals.

As deployment proceeds, the success of this Verizon AST SpaceMobile cellular service will serve as a critical test case for the integration of terrestrial and satellite networks on a commercial scale.

| Latest From Us

Picture of Faizan Ali Naqvi
Faizan Ali Naqvi

Research is my hobby and I love to learn new skills. I make sure that every piece of content that you read on this blog is easy to understand and fact checked!

This $1,600 Graphics Card Can Now Run $30,000 AI Models, Thanks to Huawei

Running the largest and most capable language models (LLMs) has historically required severe compromises due to immense memory demands. Teams often needed high-end enterprise GPUs, like NVIDIA’s A100 or H100 units, costing tens of thousands of dollars.

Table of Contents

This constraint limited deployment to large corporations or heavily funded cloud infrastructures. However, a significant development from Huawei’s Computing Systems Lab in Zurich seeks to fundamentally change this economic reality.

They introduced a new open-source technique on October 3, 2025, specifically designed to reduce these demanding memory requirements, democratizing access to powerful AI.

Key Takeaways

  • Huawei’s SINQ technique is an open-source quantization method developed in Zurich aimed at reducing LLM memory demands.
  • SINQ cuts LLM memory usage by 60–70%, allowing models requiring over 60 GB to run efficiently on setups with only 20 GB of memory.
  • This technique enables running models that previously required enterprise hardware on consumer-grade GPUs, like the single Nvidia GeForce RTX 4090.
  • The method is fast, calibration-free, and released under a permissive Apache 2.0 license for commercial use and modification.

Introducing SINQ: The Open-Source Memory Solution

Huawei’s Computing Systems Lab in Zurich developed a new open-source quantization method specifically for large language models (LLMs).

This technique, known as SINQ (Sinkhorn-Normalized Quantization), tackles the persistent challenge of high memory demands without sacrificing the necessary output quality according to the original article.

The key innovation is making the process fast, calibration-free, and straightforward to integrate into existing model workflows, drastically lowering the barrier to entry for deployment.

The Huawei research team has made the code for performing this technique publicly available on both Github and Hugging Face. Crucially, they released the code under a permissive, enterprise-friendly Apache 2.0 license.

This licensing structure allows organizations to freely take, use, modify, and deploy the resulting models commercially, empowering widespread adoption of Huawei SINQ LLM quantization across various sectors.

Shrinking LLMs: The 60–70% Memory Reduction

The primary function of the SINQ quantization method is drastically cutting down the required memory for operating large models. Depending on the specific architecture and bit-width of the model, SINQ effectively cuts memory usage by 60–70%.

This massive reduction transforms the hardware requirements necessary to run massive AI systems, enabling greater accessibility and flexibility in deployment scenarios.

For context, models that previously required over 60 GB of memory can now function efficiently on approximately 20 GB setups. This capability serves as a critical enabler, allowing teams to run large models on systems previously deemed incapable due to memory constraints.

Specifically, deployment is now feasible using a single high-end GPU or utilizing more accessible multi-GPU consumer-grade setups, thanks to this efficiency gained by Huawei SINQ LLM quantization.

Democratizing Deployment: Consumer vs. Enterprise Hardware Costs

This memory optimization directly translates into major cost savings, shifting LLM capability away from expensive enterprise-grade hardware. Previously, models often demanded high-end GPUs like NVIDIA’s A100, which costs about $19,000 for the 80GB version, or even H100 units that exceed $30,000.

Now, users can run the same models on significantly more affordable components, fundamentally changing the economics of AI deployment.

Specifically, this allows large models to run successfully on hardware such as a single Nvidia GeForce RTX 4090, which costs around $1,600.

Indeed, the cost disparity between the consumer-grade RTX 4090 and the enterprise A100 or H100 makes the adoption of large language models accessible to smaller clusters, local workstations, and consumer-grade setups previously constrained by memory the original article highlights.

These changes unlock LLM deployment across a much wider range of hardware, offering tangible economic advantages.

Cloud Infrastructure Savings and Inference Workloads

Teams relying on cloud computing infrastructure will also realize tangible savings using the results of Huawei SINQ LLM quantization. A100-based cloud instances typically cost between $3.00 and $4.50 per hour.

In contrast, 24 GB GPUs, such as the RTX 4090, are widely available on many platforms for a much lower rate, ranging from $1.00 to $1.50 per hour.

This hourly rate difference accumulates significantly over time, especially when managing extended inference workloads. The difference can add up to thousands of dollars in cost reductions.

Organizations are now capable of deploying large language models on smaller, cheaper clusters, realizing efficiencies previously unavailable due to memory constraints . These savings are critical for teams running continuous LLM operations.

Understanding Quantization and Fidelity Trade-offs

Running large models necessitates a crucial balancing act between performance and size. Neural networks typically employ floating-point numbers to represent both weights and activations.

Floating-point numbers offer flexibility because they can express a wide range of values, including very small, very large, and fractional parts, allowing the model to adjust precisely during training and inference.

Quantization provides a practical pathway to reduce memory usage by reducing the precision of the model weights. This process involves converting floating-point values into lower-precision formats, such as 8-bit integers.

Users store and compute with fewer bits, making the process faster and more memory-efficient. However, quantization often introduces the risk of losing fidelity by approximating the original floating-point values, which can introduce small errors.

This fidelity trade-off is particularly noticeable when aiming for 4-bit precision or lower, potentially sacrificing model quality.

Huawei SINQ LLM quantization specifically aims to manage this conversion carefully, ensuring reduced memory usage (60–70%) without sacrificing the critical output quality demanded by complex applications.

Conclusion

Huawei’s release of SINQ represents a significant move toward democratizing access to large language model deployment. Developed by the Computing Systems Lab in Zurich, this open-source quantization technique provides a calibration-free method to achieve memory reductions of 60–70%.

This efficiency enables models previously locked behind expensive enterprise hardware to run effectively on consumer-grade setups, like the Nvidia GeForce RTX 4090, costing around $1,600.

By slashing hardware requirements, SINQ fundamentally lowers the economic barriers for advanced AI inference workloads.

The permissive Apache 2.Furthermore, 0 license further encourages widespread commercial use and modification, promising tangible cost reductions that can amount to thousands of dollars for teams running extended inference operations in the cloud.

Therefore, this development signals a major shift, making sophisticated LLM capabilities accessible far beyond major cloud providers or high-budget research labs, thereby unlocking deployment on smaller clusters and local workstations.

| Latest From Us

Picture of Faizan Ali Naqvi
Faizan Ali Naqvi

Research is my hobby and I love to learn new skills. I make sure that every piece of content that you read on this blog is easy to understand and fact checked!

The Global AI Safety Train Leaves the Station: Is the U.S. Already Too Late?

While technology leaders in Washington race ahead with a profoundly hands-off approach toward artificial intelligence, much of the world is taking a decidedly different track. International partners are deliberately slowing innovation down to set comprehensive rules and establish regulatory regimes.

Table of Contents

This divergence creates significant hurdles for global companies, forcing them to navigate fragmented expectations and escalating compliance costs across continents.

Key Takeaways

  • While Washington champions a hands-off approach to AI, the rest of the world is proactively establishing regulatory rules and frameworks.
  • The US risks exclusion from the critical global conversation surrounding AI safety and governance due to its current regulatory stance.
  • Credo AI CEO Navrina Singh warned that the U.S. must implement tougher safety standards immediately to prevent losing the AI dominance race against China.
  • The consensus among U.S. leaders ends after agreeing that defeating China in the AI race remains a top national priority.

The Regulatory Chasm: Global AI Safety Standards

The U.S. approach to AI is currently centered on rapid innovation, maintaining a competitive edge often perceived as dependent on loose guardrails. However, the international community views the technology with greater caution, prioritizing the establishment of strict global AI safety standards.

Infographic

Companies operating worldwide face complex challenges navigating these starkly different regimes, incurring unexpected compliance costs and managing conflicting expectations as a result. This division matters immensely because the U.S.

could entirely miss out on shaping the international AI conversation and establishing future norms.

During the Axios’ AI+ DC Summit, government and tech leaders focused heavily on AI safety, regulation, and job displacement. This critical debate highlights the fundamental disagreement within the U.S. leadership regarding regulatory necessity.

While the Trump administration and some AI leaders advocate for loose guardrails to ensure American companies keep pace with foreign competitors, others demand rigorous control.

Credo AI CEO Navrina Singh has specifically warned that America risks losing the artificial intelligence race with China if the industry fails to implement tougher safety standards immediately.

US-China AI Race and Technological Dominance

Winning the AI race against China remains the primary point of consensus among U.S. government and business leaders, but their agreement stops immediately thereafter. Choices regarding U.S.-China trade today possess the power to shape the global debate surrounding the AI industry for decades.

The acceleration of innovation driven by the U.S.-China AI race is a major focus for the Trump administration, yet this focus also heightens concerns regarding necessary guardrails and the potential for widespread job layoffs.

Some experts view tangible hardware as the critical differentiator in this intense competition. Anthropic CEO Dario Amodei stated that U.S. chips may represent the country’s only remaining advantage over China in the competition for AI dominance.

White House AI adviser Sriram Krishnan echoed this sentiment, framing the AI race as a crucial “business strategy.” Krishnan measures success by tracking the market share of U.S. chips and the global usage of American AI models.

The Guardrail Debate: Speed Versus Safety

The core tension in U.S. policy revolves around the need for speed versus the implementation of mandatory safety measures, crucial for establishing effective global AI safety standards.

Importantly, many AI industry leaders, aligned with the Trump administration’s stance, advocate for minimal regulation, arguing loose guardrails guarantee American technology companies maintain a competitive edge.

Conversely, executives like Credo AI CEO Navrina Singh argue that the industry absolutely requires tougher safety standards to ensure the longevity and ethical development of the technology.

The industry needs to implement tougher safety standards or risk losing the AI race, Navrina Singh stressed during a sit-down interview at Axios’ AI+ DC Summit on Wednesday. This debate over guardrails continues to dominate discussions among policymakers.

Furthermore, the sheer pace of innovation suggests that the AI tech arc is only at the beginning of what AMD chair and CEO Lisa Su described as a “massive 10-year cycle,” making regulatory decisions now profoundly important for future development.

Political Rhetoric and Regulatory Stalls

Policymakers continue grappling with how—or whether—to regulate this rapidly evolving field at the state and federal levels. Sen.

Ted Cruz (R-Texas) confirmed that a moratorium on state-level AI regulation is still being considered, despite being omitted from the recent “one big, beautiful bill” signed into law. Cruz expressed confidence, stating, “I still think we’ll get there, and I’m working closely with the White House.”

Beyond regulatory structure, political commentary often touches on the cultural implications of AI. Rep. Ro Khanna (D-Calif.) criticized the Trump administration’s executive order concerning the prevention of “woke” AI, calling the concept ridiculous.

Khanna specifically ridiculed the directive, questioning its origin and saying, “That’s like a ‘Saturday Night’ skit… I’d respond if it wasn’t so stupid.” This political environment underscores the contentious, bifurcated nature of the AI policy discussion in Washington, as noted in the .

Job Displacement and Future Warfare Concerns

The rapid advancement of AI technology raises significant economic and security concerns, particularly regarding job displacement and the shifting landscape of modern conflict.

Anthropic CEO Dario Amodei specifically warned that AI’s ability to displace workers is advancing quickly, adding urgency to the guardrails debate. However, White House adviser Jacob Helberg maintains an optimistic, hands-off view regarding job loss.

Helberg contends that the government does not necessarily need to intervene if massive job displacement occurs. He argued that more jobs would naturally emerge, mirroring the pattern observed after the internet boom.

Helberg concluded that the notion the government must “hold the hands of every single person getting displaced actually underestimates the resourcefulness of people.” Meanwhile, Allen Control Systems co-founder Steve Simoni noted the U.S.

significantly lags behind countries like China concerning the ways drones are already reshaping contemporary warfare.

Conclusion: The Stakes of US Isolation

The U.S. Finally, insistence on a loose-guardrail approach to accelerate innovation contrasts sharply with the rest of the world’s move toward comprehensive global AI safety standards. This divergence creates significant obstacles for global companies and threatens to exclude the U.S.

from defining future international AI governance. Leaders agree on the necessity of winning the U.S.-China AI race, yet they remain deeply divided on the path to achieving that dominance, arguing over chips, safety standards, and regulation’s overall necessity.

The warnings from industry experts about the necessity of tougher safety standards—and the potential loss of the race without them—cannot be ignored.

Specifically, as the AI technology arc enters a decade-long cycle, the policy choices made in Washington regarding regulation and trade will fundamentally shape the industry’s global trajectory.

Ultimately, failure to engage with international partners on critical regulatory frameworks risks isolating the U.S. as the world pushes ahead on governance, with or without American participation.

| Latest From Us

Picture of Faizan Ali Naqvi
Faizan Ali Naqvi

Research is my hobby and I love to learn new skills. I make sure that every piece of content that you read on this blog is easy to understand and fact checked!

Don't Miss Out on AI Breakthroughs!

Advanced futuristic humanoid robot

*No spam, no sharing, no selling. Just AI updates.

Ads slowing you down? Premium members browse 70% faster.

Join Premium Today!