On March 31, 2026, Pasadena-based startup PrismML came out of stealth with 1-bit Bonsai, billed as the world’s first commercially viable end-to-end 1-bit large language model family. The flagship Bonsai 8B packs 8.2 billion parameters into just 1.15 GB of weights — a 14x size gap against comparable 16-bit models that routinely occupy 16 GB. The weights are on Hugging Face under Apache 2.0.
Why it matters: extreme quantization has long been treated as a demo-grade trick — runnable, but not useful. PrismML’s counter-thesis is that 1-bit is not post-training compression but a native architecture trained from scratch with binary weights. If the thesis holds, the tier of local model that fits on phones, robots, and wearables moves up a full class.
True 1-Bit, End to End: No Escape Hatches
Most low-bit models on the market quietly mix in high-precision components — a “1.58-bit” pipeline often keeps FP16 in the layers that matter most. PrismML stresses that Bonsai is “true 1-bit, end to end”: embeddings, attention layers, MLP layers, and the LM head are all 1-bit, with “no higher-precision escape hatches.” The family ships in three sizes — 8B (1.15 GB), 4B (0.5 GB), and 1.7B (0.24 GB) — all trained on Google TPU v4.
The whitepaper flags a longer-range bonus: if 1-bit-native hardware ever arrives, multiplications can be replaced with additions, opening another round of energy and latency gains. That is a software architecture aimed squarely at next-generation silicon.
Score vs. Density: Reading the 70.5
The honest part first: Bonsai 8B averages 70.5 across benchmarks, below Qwen3 8B’s 79.3. The line items are MMLU Redux 65.7, MuSR 50.0, GSM8K 88.0, HumanEval+ 73.8, IFEval 79.8, and BFCLv3 65.7. On absolute scores it clearly loses, and the company does not pretend otherwise.
But PrismML pitches a different yardstick: intelligence density, defined as the negative log of the error rate divided by model size. Bonsai 8B computes to 1.06/GB against 0.10/GB for Qwen3 8B — a 10.6x gap. For storage- and memory-constrained devices, that is the metric that decides what fits and what runs at all: Qwen3 8B’s FP16 weights take 16.38 GB, which simply has no seat on a phone.
Speed, Energy, and the Agentic Demo
The measured numbers: 131 tokens per second on an M4 Pro Mac, 368 on an RTX 4090, roughly 44 on an iPhone 17 Pro Max and 40 on an iPhone 17 Pro — against 23 tokens per second for a 16-bit 1B model. In other words, an 8B-class 1-bit model outruns a 1B-class 16-bit model on a flagship phone. Energy figures land at 0.074 mWh per token on M4 Pro and 0.068 mWh on iPhone 17 Pro Max, with overall efficiency claimed 4 to 5 times better.
The agentic showcase is more telling: on the same set of tool-calling tasks, Bonsai 8B completed 50 while a standard 16-bit 8B model completed 6. If that holds up, the biggest winners are offline agent apps whose bottleneck was never reasoning ability, but the fact that nothing useful fit on the device.
From Caltech Lab to Khosla’s Bet
PrismML’s technology comes out of Caltech research with exclusively licensed IP; co-founder Babak Hassibi is a Caltech professor serving as CEO. Funding comes from Khosla Ventures and Cerberus Ventures, plus compute grants from Google and Caltech — a startup betting on smaller models, bankrolled partly by the camp known for building bigger ones.
The investor framing is worth quoting. Vinod Khosla: “AI’s future will not be defined by who can build the largest datacenters.” Cerberus’s Amir Salek, formerly Google’s TPU lead, is blunter: “Power has become the ultimate bottleneck for scaling AI datacenters.” Hassibi keeps the roadmap modest: “We see 1-bit not as an endpoint, but as a starting point.”
What It Means for Developers
Three practical effects. First, deployment paths work today: MLX on Apple devices, llama.cpp’s CUDA backend on NVIDIA, and GitHub for Windows and Linux. Second, Apache 2.0 means commercial use with no strings attached — offline translation, on-device summarization, and local agent products can build directly on Bonsai without a licensing review. Third, the 70.5 average is your selection-logic reminder: this is a density-first model, and replacing cloud flagships is not in its script. The right question is how much capability you get per gigabyte, not absolute SOTA.
Sources
- Announcing 1-bit Bonsai: The First Commercially Viable 1-bit LLMs — PrismML
- PrismML Launches World’s First 1-Bit AI Model — PR Newswire
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.