PrismML released Ternary Bonsai 2 27B on September 17, offering a smaller-footprint derivative of Qwen3.8-27B under Apache 2.0. For someone with a modest local machine, the useful question is whether that smaller download leaves enough memory for the actual job. Release announcement · Checkpoint license
The release makes a local evaluation more interesting. It does not establish that every workload which runs on a larger Qwen deployment will run well on an inexpensive laptop.
Count the files you will load
The repository offers multiple representations of the same model. The PTQ1_0 language-model file is 5.95 GB; PQ2_0 is 7.21 GB. Vision adds a separate projector: 629 MB for Q8 or 931 MB for BF16. Those are download sizes, before runtime buffers and the context cache. Published files
For example, combining PTQ1_0 with the Q8 projector gives roughly 6.58 GB of files. That arithmetic is not a measured memory requirement. An 8 GB machine also has an operating system and other processes to accommodate, and the space left for a conversation depends on the runtime and configuration.
Our local LLM machine guide uses the same budgeting approach: choose the artifact, count its companions, then allow for the workload. A model’s parameter count alone is a poor purchasing specification.
The runtime is part of the choice
PrismML’s model card calls for its own llama.cpp build with custom ternary kernels. It describes CUDA, Metal and CPU paths, and links companion Apple-silicon implementations. Generic integration snippets on a model-hosting page should not override the publisher’s actual setup instructions. Model card and runtime instructions
That dependency belongs in an evaluation plan. Record the runtime version alongside the model file and prompt settings. Check whether the application you intend to use can connect to that server, and whether upgrades preserve the behavior you tested. A smaller model file is useful only if the surrounding software can run the workflow reliably.
Read the quality claim narrowly
PrismML reports retaining 98.2% of its full-precision baseline’s performance across its launch benchmark aggregate. This is a publisher comparison, not a guarantee that every task loses the same small amount of quality. Launch evaluation
We would start with representative document questions, tool calls and coding tasks that already have verifiable answers. Compare completion quality, latency, peak memory and failures at the context lengths you actually need. Include the original model in the same setup where practical. S5 Labs has not run that comparison; Bonsai 2 belongs on the evaluation list, with a clear record of what would justify adopting it.