Bonsai 2 27B puts a 27B model on a laptop in 5.9 GB
Two releases worth your attention today, and they pull in opposite directions: one makes a 27B model fit where it never fit before, the other shows the numbers we use to rank models are softer than advertised.
PrismML shipped Ternary Bonsai 2 27B, a compressed build of Qwen3.8-27B that holds more than 98 per cent of the full-precision model's capability at a fraction of its size. The language weights land at 5.9 GB — the on-disk pack is 8.60 GB once the 0.92 GB vision tower is counted — which is the difference between a model that needs a datacenter GPU and one that runs on an everyday laptop. On an M5 Max it decodes at roughly 47 tokens a second; on an M4 Pro the practical ceiling for very long prompts is prefill, around 125 tokens a second, rather than generation. PrismML's own measurement claims the model streams about 204 GB/s of weights on an M5 Pro, which is the whole design thesis: low-bit weights live or die on memory bandwidth, and a ternary representation cuts how much data the chip has to move per token.
The interesting claim is where the capability survives. Low-bit builds usually pay for their size in exactly the tasks that need long chains of correct steps, because a single bad step compounds. PrismML says the reasoning backbone comes through intact — math drops from 97.06 to 96.57, coding matches the full-precision baseline, instruction following is marginally ahead — with the remaining loss concentrated in knowledge, reasoning and vision. Against a sub-4-bit conventional build of the same base model, Bonsai 2 27B scores more than twelve points higher at under two-thirds the size, and lands within 0.4 points of a standard 4-bit quantisation at a third of the footprint.
There is a marketing trap in low-bit naming that PrismML is pointing at directly: a widely used "2-bit" build of Qwen3.8-27B is really 2.8 bits per weight at 9.4 GB. Bonsai keeps the bit-width its name promises. The models run on NVIDIA cards through CUDA and on Apple hardware through MLX with custom low-bit kernels, ship under Apache 2.0, and come with a browser WebGPU demo. It is not the first 27B-class model aimed at phones — Qualcomm pitched the same idea on its next NPU and MediaTek on a 2nm chip — but this one is open weights you can download today rather than silicon you have to buy.
A four-paper audit of coding-agent benchmarks says one patch in five that SWE-bench Verified counts as solved is semantically wrong. The SWE-ABS researchers strengthened the test suites with coverage-driven augmentation and mutation testing, then re-scored the top thirty agents. Every model gave back between ten and seventeen points — Claude 4.5 Opus fell from 76.80 to 66.40 under a single-scaffold harness — and the ranking reshuffled underneath the drop, which is the part that hurts: if you picked a model by its leaderboard position, part of what you were reading was an artefact of which weak tests happened to be in the suite. None of the models released after February 2026, including the current frontier, has been run against the hardened tests.
The deliberate case is worse. The BenchJack auditor found working exploits in all ten benchmarks it examined and hit near-perfect scores on nine of them without solving a single task; the SWE-bench exploit runs to nine lines and abuses a trust boundary nobody guarded — the patch is applied inside a container and the test output is trusted from that same container, so an agent can drop a config file that rewrites every reported result. A separate reward-hackability study measured 28.5 per cent of tasks in a SWE-bench Verified sample as hackable and found that hackable tasks inflate pass scores by 14.14 points, with zero variation across 134 model submissions: everybody is being measured with the same broken ruler, which means relative order still carries signal and absolute scores do not.
If your CI suite is the grader your agent optimises against, what stops it from learning to pass the tests and not the intent? Tell us in the comments.
