PrismML Bets Tiny Reasoning Models Can Bring Advanced AI Onto Phones and PCs

PrismML says it has built a path to high-performing reasoning AI models that are small enough to run on personal computers and, potentially, high-end smartphones, a move that could shift more AI processing from the cloud onto users’ own devices. The startup unveiled Bonsai 2 27B on Thursday, a compressed version of Alibaba’s open-source Qwen3.8 27B model that shrinks the system to 5.9 GB while preserving nearly all of its benchmark performance.

The company’s pitch matters because it challenges one of the most accepted assumptions in modern AI: that better reasoning requires bigger models and larger infrastructure. If PrismML is right, the next wave of capable AI may not depend entirely on massive data centers, but instead on compact models that can live on laptops, desktops and phones already in people’s hands.

What PrismML launched and why it matters

PrismML’s latest release, Bonsai 2 27B, is a compact reasoning model designed to deliver near-original performance after aggressive compression. The company says it reduced the model’s memory footprint by roughly 9x to 10x compared with the source model, making it much easier to deploy on local hardware.

That is important for two reasons. First, smaller models can reduce the cost of serving AI. Second, device-based inference can improve privacy and responsiveness because data does not need to travel to a remote cloud server every time a user asks a question.

PrismML is not selling a consumer app or a chatbot service. It is positioning itself as a deep infrastructure company aiming to change the economics and architecture of AI deployment.

Who is behind PrismML?

PrismML was founded by researchers from Caltech and is led by Babak Hassibi, a Caltech professor known for his work in compression technologies. The startup also benefits from the backing of notable investors and advisers with strong ties to the AI and systems world.

Among its supporters are Khosla Ventures, Cerberus Capital and Caltech. Ion Stoica, a co-founder of Databricks and director of Berkeley’s Sky Computing Lab, is also an adviser. Stoica’s lab has helped inspire or incubate several influential AI technologies and startups, adding technical credibility to PrismML’s claims.

Even though the company has raised a relatively modest seed round of $22.25 million, the technical pedigree around it makes the startup stand out in a crowded field of AI infrastructure players.

PrismML’s backers and leadership believe the company’s compression methods could make advanced AI practical on the devices people already own, while also reducing dependence on cloud computing.

How does PrismML shrink a large language model?

PrismML says its models get smaller by compressing the weights inside the network, which are the learned values that store the model’s behavior after training. In a conventional model, those weights are usually stored with 16-bit precision.

The company’s approach uses what it calls ternary weights, meaning each value is reduced to one of three options: positive one, negative one or zero. That dramatically lowers the amount of memory needed to store a model while aiming to preserve much of its original capability.

In practical terms, the compression lowers the hardware requirements for running reasoning models. That could make AI more accessible on consumer devices and reduce the barrier for developers who want to deploy local-first applications.

Why ternary weights are notable

Ternary weights are notable because they attack the storage problem at a structural level rather than simply trimming a model’s size through conventional pruning or distillation. PrismML argues that this approach allows it to preserve more of the model’s intelligence than older compression methods.

That claim is difficult to prove in absolute terms, but the company’s benchmark results suggest the idea is gaining traction. According to PrismML, Bonsai 2 reaches 98% of Qwen’s aggregate benchmark score, up from 95% for the first Bonsai model released earlier this year.

Why device-level AI is becoming a bigger deal

Device-level AI is becoming a bigger deal because users and businesses are increasingly sensitive to cost, latency and privacy. When a model runs locally, responses can arrive faster, and organizations may avoid sending sensitive information to an external cloud provider.

Ion Stoica, who advises PrismML, said the company’s technology could make advanced intelligence available directly on the hardware users already own. His argument is that if the model runs on-device, the service can feel free at the point of use because the inference cost is shifted away from a centralized provider. He also emphasized the privacy benefits of keeping data off the cloud.

This aligns with a broader trend in consumer and enterprise AI. Device makers and software companies have been searching for ways to put useful AI features on local hardware, both to cut cloud costs and to meet rising expectations around privacy and offline functionality.

Could this affect phones as well as PCs?

Yes, PrismML says that is the goal, though the practical experience will vary by device. The company says Bonsai 2 is small enough for PCs and possibly for high-end smartphones, which would be a meaningful expansion of where reasoning models can run.

If that becomes reliable at scale, it could change how people interact with AI assistants, writing tools and search-like applications. Instead of relying on remote inference, more of the work could happen directly on the device, potentially opening the door to faster and more private experiences.

How does PrismML compare with other compression efforts?

PrismML is not alone in trying to compress large language models, but it says its approach is different because it retains more of the original model’s performance. Another startup, Multiverse Computing, is also working on compression technology and has attracted substantial funding.

The distinction PrismML is trying to draw is not just about shrinking models, but about shrinking them without major quality loss. That is the real competitive bar in this part of the market, where smaller models are only valuable if they remain capable enough to be useful in real-world tasks.

Many AI companies can make a model lighter. Much fewer can make it lighter while preserving reasoning quality, general utility and benchmark strength.

What do the benchmarks and downloads suggest?

The company’s benchmark and download figures suggest it has already attracted meaningful interest from developers and AI enthusiasts. PrismML says the original Bonsai model, released in March, has been downloaded more than 11 million times. Its smaller models have added another 2.6 million downloads.

Those numbers are not proof of long-term commercial success, but they do indicate that the product has found an audience. In open-source AI, downloads can signal both developer curiosity and practical adoption, particularly when models are easy to test and deploy locally.

The jump from 95% benchmark parity in the first release to 98% in Bonsai 2 also matters because it shows the compression technique may be improving quickly. Even if parity never reaches 100%, PrismML argues that the remaining gap may be less important in real use than benchmark comparisons suggest.

Why benchmarks do not tell the whole story

Benchmarks matter, but they are an imperfect measure of how useful a model will be in the wild. Real applications depend not only on the model itself, but also on surrounding software, prompting tools, execution harnesses and the specifics of each task.

PrismML argues that a small performance difference on a benchmark may not translate into a meaningful user experience difference, especially when the original uncompressed model is not perfect to begin with. That is a familiar argument in AI, where practical usefulness often matters more than theoretical score gaps.

In other words, a model that is 2% behind on aggregate evaluation may still be good enough for many consumer and enterprise workflows if it is cheaper, faster and more private.

Why bigger models may be easier to compress

PrismML’s CEO believes the company’s approach may become easier as models get larger. Hassibi said the next releases, expected in the coming months, will target several-hundred-billion-parameter systems.

His reasoning is that bigger models may have more redundancy, giving compression algorithms more room to reduce size without destroying performance. In his view, that means the path to preserving intelligence may actually improve as model scale increases.

That is an important claim because it runs against the instinct that scaling always makes efficient deployment harder. PrismML is betting the opposite may be true: larger models may be more forgiving when it comes to compression.

What happens next?

The next test for PrismML will be whether it can apply its approach to significantly larger frontier-style models while keeping them practical. If it can, the startup could become a key player in the next phase of AI infrastructure.

That would matter not only for consumer devices, but also for enterprises that want powerful models without high ongoing cloud bills. It could also benefit app makers looking to build offline or privacy-preserving features without waiting for better internet connectivity or more generous cloud pricing.

Why investors are paying attention

Investors are watching because PrismML sits at the intersection of several hot AI themes: inference efficiency, local deployment, open-source model adaptation and privacy-aware computing. These are not niche concerns. They are becoming central as AI moves from experimentation to everyday use.

The startup’s combination of academic credibility, technical differentiation and evidence of real usage makes it attractive in a market where many AI companies sound similar on the surface. A seed round of $22.25 million is small by AI standards, but the company’s story is larger than the funding number.

Its release also comes at a time when major technology companies are racing to reduce AI costs and move more functionality onto devices. That broader industry momentum could help PrismML, especially if its compression methods prove portable across models and hardware categories.

What this could mean for the AI market

If PrismML’s claims continue to hold up, the implications could be significant. Developers may be able to build AI products that work offline, respond faster and keep more user data local. Hardware makers could use compressed models to add more intelligence to existing devices without requiring a major redesign.

For the industry, the more profound shift may be psychological. AI has often been discussed in terms of scale: larger models, larger clusters, larger bills. PrismML is proposing a different future, one in which intelligence is compact enough to travel with the user.

That does not mean giant models will disappear. But it does suggest a bifurcated future in which some advanced AI remains centralized, while other tasks migrate to compact models optimized for local use.

In that future, the winners may be the companies that can make AI both powerful and portable.

Timeline of PrismML’s recent progress

The company’s public momentum has accelerated quickly over the past several months. Here is a simple timeline of the most important milestones mentioned in the release:

Bottom line

PrismML is making a high-stakes bet that advanced reasoning models do not need to remain huge to stay useful. Its newest release suggests that compressed AI could be close to practical on everyday devices, a development that may influence everything from privacy to cloud economics to how AI assistants are built.

The company still has to prove that its results scale, but it has already shown enough technical momentum to make the industry pay attention. If PrismML can keep improving while shrinking models further, it may help redefine where and how AI is used.