Ternary Bonsai 2 27B reduces the language-model weight footprint by about 89% and scores 83.9 versus 85.4 for full-precision Qwen3.8 27B in PrismML's benchmark suite. The 5.9GB figure solves the weight-loading problem; context memory, throughput and task-level accuracy determine the usable workload.
PrismML's Ternary Bonsai 2 27B compresses the language weights of Qwen3.8 27B from roughly 54GB in FP16 to 5.9GB. The company uses ternary {-1, 0, +1} weights with FP16 group-wise scaling and reports 1.76 effective bits per weight. On its published 20-benchmark suite, the model scores 83.9 versus 85.4 for the full-precision base, which PrismML describes as 98.2% retention of aggregate benchmark performance.
The hardware arithmetic is straightforward; the quality claim needs more care. A 5.9GB weight file leaves about 10.1GB of nominal memory headroom on a 16GB GPU before accounting for the vision component, activations, runtime buffers and key-value cache. For teams that already own 16GB cards, that removes one hardware-spend trigger: a larger-VRAM accelerator is no longer required simply to load the 27B language weights. CTOL's developer testing has run the model in high-reasoning modes on 16GB cards, including the RTX 5060 Ti, for coding and 3D-modeling work.
That means 27B-class reasoning is no longer automatically excluded by the card's weight capacity. The bottleneck moves to the memory and compute generated by the actual workload.
The same footprint now carries a stronger model
PrismML's first 27B Bonsai release in July also put a Qwen3.6 27B derivative into a 5.9GB ternary footprint. The company reported roughly 95% aggregate retention against the full-precision base. Bonsai 2 keeps the deployment point and changes the quality density: Qwen3.8 is a stronger base, and PrismML reports 98.2% aggregate retention.
Its category results are close in several important areas. Coding scores 81.58 for Bonsai 2 against 82.17 for Qwen3.8 27B; math is 96.57 versus 97.06; agentic and tool-calling tasks are 77.57 versus 79.74. PrismML also reports throughput of up to 143 tokens per second on an RTX 5090 and 46.8 tokens per second on an M5 Max.
Those are vendor measurements, and the 98.2% figure is an average across PrismML's suite rather than a probability that an enterprise workload will lose only 1.8% of its quality. The published table does not provide confidence intervals or show how errors are distributed across individual tasks. A small aggregate gap can hide a larger regression on the one task that matters to a buyer.
There is still useful independent evidence from the first Bonsai generation. ThakiCloud loaded PrismML's released low-bit weights and found near-lossless perplexity behavior on a fixed passage, while straightforward public quantization methods did not reproduce the same compression point. Its review also concluded that the conversion method itself was not disclosed and was described in PrismML material as proprietary Caltech intellectual property. Independent users can test the released artifact. Reproducing the conversion process is harder because the method itself is not public.
Runtime memory becomes the constraint after the model loads
The headline number covers weights. Long-context inference adds a KV cache whose memory grows with sequence length; batching adds more; multimodal workloads add their own components. PrismML advertises a 262K-token context window, but the useful question for a workstation buyer is how much context can be sustained at the latency and throughput the application requires.
The low-bit format can also shift work from memory to compute. PrismML itself notes that native kernels matter because unpacking ternary weights can impose arithmetic overhead on some GPU architectures. A 27B model that fits easily may still be too slow for an interactive agent if each step expands context, calls tools and retries.
This is why the 16GB threshold matters most for teams whose previous blocker was VRAM. It lowers the minimum hardware needed to test local coding, private document analysis and other sensitive workloads without renting larger accelerators. The advantage narrows when latency, energy, context-cache memory or concurrency dominates the cost.
Bonsai 2 changes one part of the local-versus-cloud calculation decisively: its 27B language weights no longer force buyers above a 16GB GPU. Whether that hardware saving survives in service depends on peak memory, tokens per second, latency and task accuracy at the context lengths the workload actually uses.
Sources
CTOL direct developer testing, September 21, 2026. PrismML, "Introducing Bonsai 2 27B," September 17, 2026. PrismML Ternary Bonsai 2 27B model materials. ThakiCloud, independent review and reproduction work on the first Bonsai 27B generation.