“"A 27B-class model that fits on a phone changes what 'local AI' can mean." — PrismML launch announcement”
You know that feeling when you want an AI model good enough for coding and reasoning, but running it means either paying per-token to a cloud API or buying a GPU with 54GB of memory just to load the weights? Full-precision 27B-class models simply don't fit in the memory you have on a laptop or a phone. PrismML built Bonsai 27B to close that gap: the same Qwen3.6-27B model compressed to 5.9GB so it loads on consumer hardware. The catch, according to Hacker News testers, is that the compression cost shows up exactly where you'd notice it — in real tasks like SQL and vision, not just in the benchmark scores PrismML published.
Start with Qwen3.6-27B, a 27.8-billion-parameter model that normally needs 54GB of memory in full precision (FP16). PrismML keeps the same architecture — the same attention layers, the same MLPs — but rewrites every weight using either 1 bit (just -1 or +1) or 1.58 bits ("ternary": -1, 0, or +1) instead of 16 bits. Think of it like converting a photo from full color to pure black-and-white: you lose information, but the shapes are still there, and the file is a fraction of the size. That shrinks the model to 3.9GB (1-bit) or 5.9GB (ternary), which you can then run through llama.cpp or MLX on a normal laptop or phone instead of a data-center accelerator. The vision portion is compressed less aggressively (4-bit) because image understanding degrades faster under extreme compression — the benchmarks back this up, with vision scoring 65.19 versus 93.40 for math.
If you're building a local-first coding or agent tool and want a reasoning model that fits in under 6GB without sending code to a cloud provider, Bonsai 27B's 1-bit variant is worth testing. If you already run Gemma 4 12B QAT and it handles your workload, Hacker News testers who compared the two directly found Gemma held up better on practical tasks like SQL — so switching isn't automatically an upgrade. It's not for you yet if you need reliable agentic tool-calling or vision accuracy close to the FP16 baseline; both categories show the steepest score drops in PrismML's own benchmarks.
It's experimental, not production-ready: PrismML's own model card admits native ternary hardware kernels don't exist yet, so the model actually runs in 2-bit inference slots, and agentic coding "is not yet optimized in this release." Hacker News testers corroborate this with real reasoning-loop and tool-calling problems, and no independent lab has reproduced PrismML's benchmark numbers. Worth a weekend test if you're curious about the compression technique itself, not worth betting a product launch on yet.
Deep-dive insight, Easy and Pro modes, plus action playbooks — the full breakdown is one tap away.