PrismML Releases Bonsai 2, a Compact Reasoning Model That Runs on PCs

U.S. AI research startup PrismML on September 17, 2026, released the reasoning language model “Bonsai 2 27B.” The model is a compressed derivative of Alibaba’s open model “Qwen3.8 27B.” Its size is kept to 5.9GB, intended for execution on personal computers. It is designed to complete high-performance reasoning processing entirely on-device. In reporting by TechCrunch AI’s Julie Bort, the depth of the technical team and the novelty of the approach were highlighted as points of interest.

If AI lab PrismML isn’t on your radar yet, it should be — not because it’s raised gobs of money (it hasn’t yet, just a $22.25 million seed round), but because of the technical minds involved and the potentially industry-changing tech it’s developing.

The company remains small, with only a seed round in the $22 million range. Nevertheless, it has made its presence felt with technology for miniaturizing large reasoning models. Regarding rumors of talks with Apple, Chief Executive Officer Babak Hassibi reportedly declined to answer. The move suggests strong industry interest in on-device deployment.

Overview of Bonsai 2 Compressed to 5.9GB

Bonsai 2 27B is a compressed model based on Qwen3.8 27B. Memory usage is reduced by 9 to 10 times compared to the original. At 5.9GB, it is small enough to fit on current personal computers. Operation on high-performance smartphones is also in view. The aim is to open up reasoning applications that do not assume cloud connectivity.

A reasoning model refers to the type that performs step-by-step deliberation before answering. They have traditionally required large amounts of memory and computing resources. Execution was therefore limited to data-center-scale computing infrastructure. Bonsai 2 is an attempt to ease this constraint. By keeping inference on-device, reduced latency and communication costs can be expected.

The company released the first-generation Bonsai in March. The first generation reached 95% of the original’s score in aggregate. The company says Bonsai 2 improves that to 98%. This shows improved compression accuracy between generations. Because the original model is highly open, third-party verification is expected to proceed.

How Ternary-Weight Compression Technology Works

At the core of language models are groups of numbers called “weights.” Weights serve to retain knowledge acquired during training. They are normally recorded at 16 bits each. In large-scale models, the total becomes enormous, straining onboard memory. The success of compression depends on this representational efficiency.

PrismML’s method is called “ternary” weights. It is a technique that simplifies values to three types: +1, -1, and 0. The amount of information to be retained is greatly reduced, lowering size and transfer load. Simplified computation also curbs power consumption during execution. It has properties well suited to on-device deployment.

The company cites almost no performance degradation as its strength. The basis is its claim of maintaining 98% in aggregate scores. The remaining 2% gap is often meaningless in practical use. That is because the scores themselves do not fully reflect real-world accuracy. Tuning of the surrounding programs that run the model also affects accuracy.

The first-generation model is said to have exceeded 11 million cumulative downloads. Its smaller derivatives also saw 2.6 million downloads. Developer interest is reflected in the numbers. It is at a stage where distribution as a public resource is progressing. Confirming reproducibility will be a future challenge.

Technical Background Seen in Founding Team

and Support Structure

PrismML originates from a group of researchers at the California Institute of Technology. It is led by Hassibi, a professor there and an expert in compression technology. Adviser Ion Stoica is among its advisers. He is known as a co-founder of Databricks. He also heads Berkeley’s Sky Computing Lab.

That lab has produced technologies and startups such as Letta and SGLang. It is a place that gathers talent strong in surrounding infrastructure for language models. It has high affinity with PrismML’s direction. It is an organization positioned at the intersection of academia and industry. Its depth of talent underpins the reliability of its technology.

On the funding side, Khosla Ventures is an investor. Cerberus Capital and the California Institute of Technology are also listed. The amount raised remains at the seed stage at $22.25 million. It is advancing research and development without relying on large-scale fundraising. It is attracting attention as a cost-effective operation.

Rumors of talks with Apple remain unconfirmed. Hassibi declined to answer when asked. Collaboration with device manufacturers is directly linked to expanding the uses of small models. Whether any partnership announcements follow will be the focus. It should be noted that at this point it remains speculation.

Competitive Trends and Perspectives for

Evaluating Performance Retention

PrismML is not the only company working on compressing language models. Multiverse Computing is cited as a representative competitor. It was founded by a professor at the Donostia International Physics Center. It has carried out large fundraising and is expanding its business. Competition over the practical application of compression technology is intensifying.

Hassibi places his company’s uniqueness in its performance retention rate. He says the ability to maintain results almost equivalent to the original is a differentiating factor. The figure of 98% is a claim based on aggregate scores. It may vary depending on measurement conditions and task selection. Independent re-evaluation is needed for corroboration.

Whether 100% parity can be reached remains uncertain. He himself indicates that some impact from compression remains. Complete agreement is likely to remain of only academic interest. That is because errors in the original itself and biases in scores exist. The 2% gap can be interpreted as having limited impact in practice.

The importance of surrounding programs has also been pointed out. Even with the same model, output accuracy varies with the execution framework. The design of retrieval integration and procedure control affects results. Evaluation of compressed models is not complete with standalone scores alone. Verification across the entire operation becomes essential.

Changes in Usage Patterns Brought by

On-Device Execution

Reasoning models that run on-device change usage patterns. Advanced text processing becomes possible even in environments without connectivity. There is the advantage of processing confidential documents without sending them externally. It also leads to shorter response latency. In terms of cost structure, outsourcing of communication and computation can be reduced.

PC deployment is the first realistic stage. At 5.9GB, it fits within current memory capacities. Expansion to high-performance mobile devices will be the next stage. Memory and heat constraints on mobile devices remain as challenges. Optimization and semiconductor advances will determine the speed of adoption.

Opening of on-device graphics execution infrastructure is also progressing. As Mesa Rusticl and Mali Panfrost Enabled by Default shows, development of open execution environments continues. The spread of small models is linked to such infrastructure development. Developers can test without being tied to specific closed environments. Expanded choices will be a factor in broadening the range of applications.

For enterprise use, private deployment and audit compliance become issues. Designs that curb external transmission tend to be viewed favorably by administrative departments. On the other hand, operations for model updates and vulnerability response are needed. Distribution and version management mechanisms become important. Accumulation of operational knowledge for device-distributed operation is required.

Editorial Opinion

On short-term impact, in the next 3 to 6 months, reproduction and verification by developers is seen to become active. The focus will be on which tasks the aggregate figure of 98% holds for. Reports on combinations with execution frameworks are also seen to increase. It is assessed that reports of operation on mobile devices will be shared, clarifying practical conditions.

On the long-term view, in 1 to 3 years, on-device inference could grow into a standard option. Designs that do not assume connectivity are assessed to deliver value in confidential processing and offline environments. Whether there is collaboration with specific companies is seen to affect the speed of mobile device adoption. If miniaturization of open models progresses, leadership in applications could be said to become distributed.

Questions from the editorial team are presented. It is asked in which operations the remaining 2% gap is meaningful. It is also asked whether the aggregation method of scores reflects real-world failures. It is seen that whether the large number of downloads indicates continued use also needs verification. Dependence of the business foundation on research organizations is also assessed as an issue.

References

• “PrismML hopes its tiny LLM will change how we all use AI”, by Julie Bort — TechCrunch AI, 2026-09-17T22:34:09.000Z (ARR)
• Source URL: https://techcrunch.com/2026/09/17/prismml-hopes-its-tiny-llm-could-change-how-we-all-use-ai/