PrismML hopes tiny models can do work people now send to the cloud, and on Thursday, September 17, 2026, the Caltech spinout released Bonsai 2 27B to make the case. The open-weight model squeezes Alibaba’s Qwen3.8 27B down to 5.9 GB, and PrismML says it keeps 98.2% of the original’s benchmark score.
That figure is PrismML’s, not an independent one. The company ran its 20-benchmark suite itself, in thinking mode, and TechCrunch‘s Julie Bort relayed the numbers on September 17 alongside comments from CEO Babak Hassibi. As far as I can find, no third party has published a run of Bonsai 2 27B yet.
Here is the split. PrismML’s own announcement documents the file size, the license and the scores, though the scores are its own measurements. The claim that the compression is near lossless is a company claim. Hassibi’s plan to apply the method to models with several hundred billion parameters, which he hopes to release within a couple of months per TechCrunch, is forward-looking and unverified. The details, and several places where the coverage does not line up, are below.
By Alex Carter. Published September 18, 2026. The news broke September 17, 2026.
How Did PrismML Get to Bonsai 2?
PrismML came out of stealth on March 31, 2026, according to its launch announcement. The first PrismML Bonsai family used 1-bit weights and was led by an 8B model that, per HPCwire, needs roughly 1 GB of memory against about 16 GB for a typical 16-bit equivalent. HPCwire also relayed the company’s claims of up to eight times faster processing and 75 to 80% lower energy use on existing hardware.
The Bonsai 8B benchmark figures at launch came from PrismML itself. A Forbes contributor’s writeup put the company’s average at 70.5 across six benchmark categories. Ternary versions followed in April, using weights of -1, 0 or +1 and trading some size for quality. Ternary Bonsai 8B is 1.75 GB against 1.15 GB for the 1-bit 8B, according to PrismML’s Hugging Face model card.
On July 14, PrismML released Bonsai 27B, built on Alibaba’s Qwen3.6 27B. It came as a 1-bit build of about 3.9 GB aimed at the iPhone 17 Pro, plus a ternary build that Quartz reported kept more than 95% of full-precision benchmark performance, per the company. The same day, Hassibi told CNBC that Apple and other companies were evaluating the models.
What PrismML Hopes Tiny Models Will Do for Local AI
PrismML’s announcement frames the goal as more useful intelligence per unit of memory, compute and power. It points to local coding agents, computer use, private document analysis, and hybrid setups where a local model handles sensitive or frequent tasks and escalates selectively to the cloud.
What Is Inside Bonsai 2 27B?
Bonsai 2 27B stores each weight as -1, 0 or +1 with FP16 group-wise scaling, which PrismML puts at 1.76 effective bits per weight. Standard weights take 16 bits each, TechCrunch(https://techcrunch.com/2026/09/17/prismml-hopes-its-tiny-llm-could-change-how-we-all-use-ai/)(https://techcrunch.com/2026/09/17/prismml-hopes-its-tiny-llm-could-change-how-we-all-use-ai/) notes. The model has 27.8 billion parameters, a 262K-token context window, text and image input, and an Apache 2.0 license. PrismML’s press release says it was trained on Google v5 TPUs.
PrismML reports up to 143 tokens per second on an NVIDIA GeForce RTX 5090 and 46.8 tokens per second on an M5 Max. On an RTX 4090 it measured 0.714 mWh per token, which the company says is 40% more energy efficient than a full-precision 8B model. All of these are company measurements.
How Do the Benchmark Numbers Break Down?
PrismML’s aggregate is 83.9 for Bonsai 2 against 85.4 for full-precision Qwen3.8 27B, which is where the 98.2% comes from. The scores below are PrismML’s, run in thinking mode. The gap column is my arithmetic.
The loss is not evenly spread. Vision, knowledge and reasoning, and agentic tool calling carry the largest gaps, at 3.05, 2.71 and 2.17 points. Coding and math are within about half a point, and instruction following scores above the full-precision model.
My read: the more useful comparison is with the previous generation. Bonsai 2’s 83.9 edges full-precision Qwen3.6 27B at 83.6, but it trails that older model on agentic tool use, coding, knowledge and reasoning, and vision. So the fair summary is roughly last generation’s full-precision quality at about a ninth of the size, not nothing lost. That is still a strong result.
Can You Run It Today?
The weights are out now. PrismML points developers to its Hugging Face collection and its Bonsai-demo repository on GitHub, which showed about 2,300 stars and 244 forks when I checked on September 18.
The 5.9 GB figure is not the whole memory bill. PrismML’s docs say the vision tower ships separately, at 0.63 GB in the 4-bit version the demo uses. PrismML’s Hugging Face page for the first ternary 27B adds that a device must hold peak memory, meaning weights plus the KV cache, activations and runtime buffers, with about 1.3 GB cited across backends. In the independent test covered below, the earlier ternary model needed 12.9 GiB, with a 4-bit KV cache, to hold its full 262K context.
Ollama users should check before assuming. A community repack of Ternary Bonsai 8B says PrismML’s official Q2_0 files loaded only in PrismML’s llama.cpp fork as of June, and it offers a version for stock llama.cpp, LM Studio and Ollama. PrismML’s repo now says its earlier formats are landing in mainline llama.cpp. I found no official Ollama listing for Bonsai 2 27B.
Where Do the PrismML Claims and Reality Diverge?
Start with who ran the tests. PrismML did, and as far as I can find, no third party has published a run of Bonsai 2 27B as of September 18. PrismML’s post calls the result practically “lossless”, which is a company claim, and the 98.2% aggregate hides the category gaps above. In fifteen years of covering launches, I have watched near lossless claims settle into lossless on the vendor’s benchmark mix until outsiders test real workloads.
The closest outside data covers the earlier model. An independent GitHub benchmark ran Ternary Bonsai 27B against a differently compressed Qwen3.6 27B build on one 16 GB RTX 5060 Ti. MMLU-Redux was a statistical tie at 0.871 against 0.860. On AIME26 with a 60,000-token thinking budget, Bonsai scored 0.867 against 0.633, but at 30 problems the author calls that accuracy gap marginal (p=0.072) and says the margin depends heavily on the thinking budget. The test never compared against full precision, so it does not verify the 95% or 98.2% figures. Its author also says the analysis was AI-orchestrated and human-reviewed.
The money figures do not line up. TechCrunch reports a $22.25 million seed round. At the March 31 launch, HPCwire reported $16.25 million, and Dealroom described it as a SAFE and seed round. I found no public explanation for the $6 million gap, so treat $22.25 million as TechCrunch’s number until PrismML clarifies. PrismML’s Bonsai 2 post also names Google and Samsung as supporters, which TechCrunch’s list of backers does not.
TechCrunch calls the Apple talks rumored and says Hassibi declined to comment. That undersells what is on the record. Apple did not comment. No deal, terms or product plans are confirmed, and one aggregator’s acquisition headline goes beyond what CNBC reported.
Two more corrections to the record. TechCrunch dates the 95% model to March, but PrismML’s own posts put the first Bonsai family on March 31 and the first Bonsai 27B on July 14, so I use those dates. And TechCrunch says the 5.9 GB model could possibly fit a high-end phone. PrismML’s July phone number, 11 tokens per second on an iPhone 17 Pro, was for the roughly 3.9 GB 1-bit build of the earlier 27B. Bonsai 2 27B launched as a ternary model, its announced speeds are for an RTX 5090 and an M5 Max, and I found no phone figure. TechCrunch also relays PrismML’s claim of more than 11 million downloads for the original Bonsai and 2.6 million for smaller models. Those counts come from the company, and I could not check them.
One detail standard coverage skipped: Quartz reported that Caltech holds the relevant patents and grants PrismML an exclusive license to commercialize them. Apache 2.0 covers the released weights, so anyone planning to reimplement the method rather than use the weights should read the patent position first.
Who Should Care About Bonsai 2 27B?
Developers running coding agents locally are the obvious audience. PrismML’s demos show Cline coding and computer-use workflows on an RTX 5090, but demos are not benchmarks, and PrismML’s own post notes that small errors compound over long agent tasks. The 2.17-point tool-calling gap is the number to watch there.
Companies with privacy or cost limits get a hybrid option: keep sensitive or high-volume work on local hardware and send hard cases to a cloud model. Ion Stoica, the Databricks co-founder who advises PrismML, told TechCrunch the appeal is intelligence that runs on a device people already own and never touches the cloud. Cloud capacity is a live constraint, as when OpenAI paused ChatGPT Pro sign-ups after Astra demand strained its systems. In my view, a 27B model complements frontier models rather than replacing them.
PrismML is also downstream of Alibaba. Every Bonsai 27B model so far is built on a Qwen base, and Quartz reported PrismML saying in July that Google’s Gemma is next. That ties it to the open-weight supply that enterprises weighing cheaper Chinese models are already sorting through, as BriefFlash has reported. Everyday users should wait, since there is no published phone speed for Bonsai 2 27B.
What Happens Next?
The next milestone is scale. Hassibi told TechCrunch that PrismML hopes to release models in the several-hundred-billion-parameter range within a couple of months. If that holds, my estimate is mid-November to December 2026. His argument is that bigger models have more room to compress, and that for larger models “it’s easier to get to 100%.” That is a forward-looking claim with no public evidence at that scale yet.
Three things would confirm the story: independent runs of Bonsai 2 27B that land near 98% on outside workloads, a published phone speed, and a clear account of PrismML’s funding. Independent results showing steep losses in long-horizon agent tasks, or a larger release that retains less than the 27B did, would undercut it. Any Apple statement would change the commercial picture, though none is confirmed.
PrismML’s March launch also framed the technology as helping datacenter buildouts do more with fewer resources and avoid ballooning energy costs. That lands differently as communities push back on new sites, such as the data center backlash in Philadelphia.
If any fact here changes, BriefFlash will add an Updated line with the date and what changed.
Key Takeaways
- PrismML hopes tiny models can take over some cloud work: on September 17, 2026 it released Bonsai 2 27B, a 5.9 GB ternary build of Alibaba’s Qwen3.8 27B under Apache 2.0.
- PrismML reports an aggregate score of 83.9 against 85.4 for full-precision Qwen3.8 27B, or 98.2%. These are company-run results, and I found no independent test of Bonsai 2 yet.
- The gap is largest in vision (3.05 points), knowledge and reasoning (2.71) and agentic tool calling (2.17), while instruction following scores above the full-precision model.
- Coverage conflicts on funding ($22.25 million per TechCrunch, $16.25 million at the March launch) and on Apple, where the CEO told CNBC in July that Apple is evaluating the models. No deal is confirmed.
FAQ
What is Bonsai 27B?
Bonsai 27B is PrismML’s first 27B-class model family, announced July 14, 2026. It is built on Alibaba’s open-source Qwen3.6 27B rather than a from-scratch base, and it shipped in two versions: a 1-bit build of about 3.9 GB aimed at phones like the iPhone 17 Pro, and a ternary build that PrismML says kept more than 95% of full-precision benchmark performance. Both are Apache 2.0. On September 17, 2026, PrismML followed with Ternary Bonsai 2 27B, based on Qwen3.8 27B, which is 5.9 GB and retains 98.2% by PrismML’s own testing.
Who is PrismML?
PrismML is a Pasadena, California AI lab spun out of Caltech that compresses open-source language models so they can run on PCs and phones. It is led by CEO Babak Hassibi, a Caltech professor and compression expert, and was co-founded with Sahin Lale, Omead Pooladzandi and Reza Sadri. Databricks co-founder Ion Stoica is an adviser. The company came out of stealth on March 31, 2026. TechCrunch lists Khosla Ventures, Cerberus Capital and Caltech as backers and reports a $22.25 million seed round, while March launch coverage reported $16.25 million, a gap PrismML has not publicly explained as far as I can find.
What is one bit Bonsai?
1-bit Bonsai is PrismML’s first model family, released March 31, 2026 under Apache 2.0. Each weight is stored as a single bit that maps to a negative or positive scale value, with every group of 128 weights sharing one FP16 scale, which works out to about 1.125 effective bits per weight according to The Machine Herald. The flagship 8B model packs about 8.2 billion parameters into 1.15 GB, against 16.38 GB at FP16. PrismML claims up to eight times faster processing and 75 to 80% lower energy use, and those figures are the company’s own. A 1-bit Bonsai 27B of about 3.9 GB followed on July 14, 2026.
What is ternary Bonsai?
Ternary Bonsai is PrismML’s second format. Instead of one bit per weight, each weight takes one of three values, -1, 0 or +1, which is about 1.58 bits in theory. The family arrived in April 2026 at 8B, 4B and 1.7B sizes, and PrismML’s model card lists Ternary Bonsai 8B at 1.75 GB against 1.15 GB for the 1-bit 8B, trading some size for quality. With FP16 group scaling, Bonsai 2 27B works out to 1.76 effective bits per weight and 5.9 GB, and it is released as a ternary model.
