The landscape of artificial intelligence is currently undergoing a paradigm shift, driven largely by the pursuit of massive computational infrastructure, sprawling data centers, and billion-dollar training budgets. Yet, while major tech conglomerates pour unprecedented capital into scaling up foundational models, a stealthy Caltech-born startup named PrismML is betting on the exact opposite approach. The company is proving that highly capable, high-performing reasoning large language models do not need to be massive to be effective. Instead, PrismML is pioneering advanced compression techniques capable of shrinking sophisticated reasoning models down to a fraction of their original size—small enough to run locally on standard personal computers and high-end smartphones.
This technical achievement is spearheaded by a modest $22.25 million seed funding round and an elite roster of academic and industry heavyweights. Led by Babak Hassibi, a professor at the California Institute of Technology (Caltech) and a recognized expert in signal processing and compression technologies, PrismML is rapidly gaining traction across the tech sector. The startup also benefits from the guidance of Ion Stoica, a co-founder of Databricks and the director of University of California, Berkeley’s Sky Computing Lab—an incubator known for spawning prominent AI infrastructure ventures. Backed by institutional heavyweights such as Khosla Ventures, Cerberus Capital, and Caltech itself, PrismML is positioning itself at the vanguard of edge-AI deployment.
The release of Bonsai 2 27B marks a major technical milestone for the company and the broader generative AI ecosystem. By compressing Alibaba’s widely utilized open-source Qwen3.8 27B model, PrismML has managed to reduce its memory footprint to just 5.9 gigabytes. This represents an unprecedented nine- to ten-fold reduction in memory consumption compared to the original architecture, rendering a sophisticated 27-parameter model lightweight enough to operate on standard consumer hardware without relying on cloud infrastructure. Rumors surrounding potential hardware integration partnerships—including unconfirmed industry reports of ongoing dialogues with Apple regarding iPhone-optimized compression—highlight the immense commercial interest in running advanced AI natively on mobile silicon.
The Evolution of Ternary Weight Compression
At the heart of PrismML’s technical differentiation is its novel approach to model quantization and weight reduction. Traditional large language models rely on floating-point representations, typically requiring 16 bits of memory for every single weight parameter. These weights serve as the foundational numerical values that a neural network learns and stores during the intensive training phase.
PrismML disrupts this convention by employing a methodology known as "ternary" weights. This technique simplifies the numerical values associated with each weight down to just three discrete options: positive one (+1), negative one (-1), or zero (0). By drastically restricting the storage requirement for every individual parameter, the overall model size drops exponentially without fundamentally collapsing the underlying neural pathways. The underlying mechanics of this open-source compression framework are publicly documented via the project’s GitHub repository, titled ternaryLLM, inviting scrutiny and collaboration from the broader developer community.
The journey toward this level of efficiency has been iterative. PrismML introduced its foundational Bonsai model family earlier in the year, capturing the attention of developers worldwide. According to company metrics, the original Bonsai iteration—released in March—achieved a cumulative benchmark score matching 95 percent of the uncompressed Qwen baseline, accumulating over 11 million downloads. Furthermore, lighter variants developed by the startup have collectively surpassed 2.6 million downloads.
With the launch of Bonsai 2 27B, PrismML has elevated its benchmark fidelity to 98 percent parity with the original Qwen architecture. This steady upward trajectory demonstrates a closing performance gap between full-scale cloud models and their highly compressed edge counterparts.
Addressing the Performance Gap and Practical Implications
In the realm of machine learning, achieving 100 percent parity between compressed and uncompressed models has long been considered the holy grail. However, industry experts argue that chasing absolute numerical perfection on standardized benchmarks is largely an academic exercise.
CEO Babak Hassibi acknowledges that some degree of performance degradation is an inherent byproduct of aggressive compression. Nevertheless, real-world utility rarely hinges on a fractional discrepancy in benchmark evaluations. Standard uncompressed large language models frequently exhibit variances in factual accuracy, while traditional benchmarks imperfectly reflect the nuanced demands of practical workflows. Moreover, advancements in surrounding software architecture—often referred to as the "harness" in which a model executes—play a critical role in maximizing operational accuracy and output reliability, frequently compensating for minor reductions in raw model capability.
Ion Stoica emphasizes that the true value proposition of edge-optimized AI extends far beyond benchmark metrics, focusing instead on accessibility, cost-efficiency, and data privacy. Running complex reasoning models locally eliminates the recurring latency, computational costs, and subscription fees associated with cloud-based API calls.
"You are going to have intelligence at your fingertips, and it’s going to be free because it’s going to run on the device you already bought," Stoica noted. "It’s also going to be private, because you’re not going to send it to the cloud."
For enterprises and individual consumers alike, this paradigm shift addresses growing concerns regarding data sovereignty. By processing sensitive personal, financial, or corporate information entirely on-device, users mitigate the security risks inherent in transmitting proprietary data to third-party cloud servers.
Future Roadmap: Scaling Up to Scale Down
While PrismML has demonstrated significant success with mid-sized open-source architectures, the startup’s ambitions extend well beyond the 27-parameter threshold. The engineering team’s immediate roadmap focuses on applying the ternary weight compression methodology to significantly larger foundational models.
According to Hassibi, upcoming model releases scheduled for the coming months will target the several-hundred-billion-parameter range. Counterintuitively, the startup anticipates that scaling up the initial size of the model may actually simplify the compression process rather than complicate it.
"The next models that we will release, hopefully in the next couple of months, will be in the several-hundred-billion-parameter range, and I expect it will be easier to retain the intelligence there," Hassibi stated. As model size increases, the sheer redundancy embedded within the neural network provides more latitude for aggressive reduction techniques without sacrificing core cognitive capabilities. Consequently, larger architectures may prove more amenable to achieving near-complete intelligence retention post-compression.
Competitive Landscape and Market Dynamics
PrismML is not operating in an isolated vacuum. The commercial imperative to bring powerful AI models to resource-constrained hardware has catalyzed a competitive sub-sector within machine learning research. Other entities, such as Multiverse Computing—founded by prominent researchers from Spain’s Donostia International Physics Center—are likewise pursuing model compression technologies and have successfully secured substantial venture capital financing.
However, PrismML’s rapid open-source adoption, combined with its strong academic pedigree rooted in Caltech and UC Berkeley research networks, positions the startup as a formidable contender. The convergence of hardware innovation from silicon manufacturers and software compression breakthroughs from startups like PrismML suggests that the next generation of artificial intelligence will not be confined to centralized server farms.
As the industry looks toward an increasingly decentralized future, the ability to pack frontier-grade reasoning into a standard consumer device could redefine consumer expectations, shifting the locus of artificial intelligence computation from the cloud directly into the palm of the user’s hand.
