Summary: PrismML emerged from stealth on March 31, 2026 with Bonsai, a family of end-to-end 1-bit LLMs whose 8B flagship occupies 1.15 GB and ships under Apache 2.0. The company argues intelligence density fits phones and local agents even though Bonsai 8B trails Qwen3 8B on average benchmark scores, and its field results still need independent scrutiny.
On March 31, 2026, a startup headquartered in Pasadena ended its stealth phase: PrismML unveiled the Bonsai family of 1-bit LLMs — put forward as the first commercially viable, end-to-end 1-bit large language models in the world. The flagship Bonsai 8B carries 8.2 billion parameters in weights that occupy only 1.15 GB, whereas comparable 16-bit models typically demand 16 GB — a 14-fold size difference. The weights are up on Hugging Face under an Apache 2.0 license.
The significance: extreme quantization has spent years boxed in as a curiosity — something that executes but serves no real purpose. PrismML argues the opposite: 1-bit here is not compression bolted on after training; the model is natively architected and trained from the ground up on binary weights. Should that argument hold, the class of local model that fits inside phones, robots, and wearables steps up a level.
One Bit All the Way Through: Nothing Held Back at Higher Precision
Low-bit offerings in the market tend to blend in high-precision parts; pipelines advertised as 1.58-bit frequently keep FP16 wherever the work is hardest. PrismML’s claim for Bonsai is “true 1-bit, end to end” — embeddings, attention layers, MLP layers, and the LM head are each 1-bit, with “no higher-precision escape hatches.” The lineup comes in three sizes, 8B (1.15 GB), 4B (0.5 GB), and 1.7B (0.24 GB), every one of them trained on Google TPU v4.
The whitepaper also points at a payoff further out: should dedicated 1-bit silicon appear, multiplies give way to additions, and energy and latency stand to fall again. The software design, in other words, is already aimed at the next round of chips.
The 70.5 Average, and the Density Argument Wrapped Around It
Start with the uncomfortable figure: across its benchmark suite, Bonsai 8B posts an average of 70.5, trailing Qwen3 8B at 79.3. Item by item: MMLU Redux 65.7, MuSR 50.0, GSM8K 88.0, HumanEval+ 73.8, IFEval 79.8, BFCLv3 65.7. On raw scores, the round clearly goes to its rival — and the company doesn’t claim otherwise.
PrismML’s response is a second yardstick, intelligence density: take the negative log of the error rate and divide by model size. Run the math and Bonsai 8B lands at 1.06/GB while Qwen3 8B sits at 0.10/GB, a 10.6x spread. On devices where storage and memory are the binding constraints, that is the number that decides what fits and what runs: at 16.38 GB of FP16 weights, Qwen3 8B has no seat on a phone.
Measured in the Field: Throughput, Power Draw, Tool Calls
The company’s own measurements: 131 tokens per second on an M4 Pro Mac, 368 on an RTX 4090, somewhere near 44 on an iPhone 17 Pro Max, and 40 on an iPhone 17 Pro. The 16-bit 1B reference model manages 23 — meaning an 8B-scale 1-bit model beats a 1B-scale 16-bit one on a flagship handset. On power, each token costs 0.074 mWh on the M4 Pro and 0.068 mWh on the iPhone 17 Pro Max, and overall energy efficiency is claimed to improve 4 to 5 times.
The agent demo lands harder: given the same tool-calling workload, Bonsai 8B finished 50 tasks while a conventional 16-bit 8B model finished 6. Should those numbers survive scrutiny, the earliest beneficiaries are offline agent products — apps whose blocker was never reasoning quality but the impossibility of fitting a useful model on the device.
The Tech and the Money Behind It
Bonsai grows out of Caltech research, with the IP secured under an exclusive license; co-founder Babak Hassibi teaches at Caltech and runs the company as CEO. The cap table lists Khosla Ventures and Cerberus Ventures, alongside compute grants from Google and Caltech — a firm wagering on smaller models, funded in part by the crowd famous for larger ones.
The backers put the thesis in their own words. Vinod Khosla: “AI’s future will not be defined by who can build the largest datacenters.” Amir Salek of Cerberus, who previously led Google’s TPU work, goes further: “Power has become the ultimate bottleneck for scaling AI datacenters.” Hassibi keeps expectations level: “We see 1-bit not as an endpoint, but as a starting point.”
The Developer Takeaways
Three consequences worth acting on. One, deployment is a same-day exercise: MLX for Apple hardware, the CUDA backend of llama.cpp for NVIDIA, GitHub as the entry point for Windows and Linux. Two, Apache 2.0 removes conditions from commercial use — products like offline translation, on-device summarization, and local agents can be layered straight onto Bonsai with no licensing review in the way. Three, treat the 70.5 average as selection guidance: this is a density-first model, and displacing cloud flagships appears nowhere in its plan. The question to ask is how much capability each gigabyte buys you, not absolute state of the art.