PrismML released Ternary Bonsai 2 27B on September 17, 2026, compressing Qwen3.8 27B from a 53.80 GB FP16 baseline into a 5.93 GB Apache 2.0 checkpoint that retains 98.2% of the parent model's aggregate benchmark score. The compression ratio is more than 9x. On PrismML's 20-benchmark suite, the ternary build scores 83.9 against the parent's 85.4.

The mechanism is ternary quantization: every weight is mapped to −1, 0, or +1, producing an effective 1.76 bits per weight across 27.36B parameters (24.35B language backbone, 2.54B in embeddings and the LM head, 0.47B in the optional vision tower shipped as a separate 0.63 GB file). The architecture keeps 75% linear attention and 25% full attention, with a 262K-token context window intact.

Where the retention actually lands matters more than the headline average. Math drops 0.49 points (96.57 vs. 97.06). Coding drops 0.59 points (81.58 vs. 82.17). Instruction following rises 1.41 points to 82.66, edging out the full-precision baseline. AIME26 comes in at 95.83 and LiveCodeBench v6 at 90.07. That's the near-lossless story.

The losses cluster elsewhere. Knowledge and reasoning drops 2.71 points, agentic tool use drops 2.17, vision drops 3.05. Long-horizon agentic work takes the hardest hit: Terminal-Bench 2.1 falls from 69.7 to 52.8, and SWE-bench Verified falls from 80.6 to 60.8, roughly 75% retention. The medium-effort average lands at 79.3 against the FP16 baseline's 82.6.

The comparison that reframes the release is against other small builds. IQ2_XXS packs Qwen3.8 27B into 7.3 GB but averages just 75.2, with AIME26 at 78.6 and LiveCodeBench v6 at 70.05. On DataNorth's independent 14-benchmark run, Bonsai 2 lands within 0.4 points of the 4-bit UD-Q4_K_XL build at roughly a third of its 17.6 GB size.

Throughput is the other headline. PrismML reports 143 tokens/second on an RTX 5090 and 46.8 tokens/second on an Apple M5 Max, with newsroom figures citing 0.714 mWh per token on an RTX 4090. The 5.95 GB PTQ1_0 and 7.21 GB PQ2_0 packings fit inside 16 GB of unified memory on a consumer laptop.

The catch is deployment. Stock llama.cpp won't load these weights. Running Bonsai 2 27B requires PrismML's own llama.cpp fork, which means the release is gated on either upstream kernel integration or the willingness of downstream tools to ship the fork. Until then, the compression story is real and the accessibility story is contingent.

Ternary quantization has been the theoretical target of small-model research for two years. This is the first release where the retention numbers on math and coding survive contact with a serious benchmark suite. The rest is a packaging problem.

Sources

https://prismml.com/news/bonsai-2-27b
https://www.prnewswire.com/news-releases/prismml-launches-bonsai-2-27b-its-most-capable-model-yet-302882228.html
https://www.marktechpost.com/2026/09/18/prismml-releases-ternary-bonsai-2-27b-a-5-9-gb-apache-2-0-model-retaining-98-2-of-qwen3-8-27b-performance/
https://alphasignal.ai/news/prismml-squeezes-qwen3-8-27b-into-5-9-gb-with-98-performance-retained
https://datanorth.ai/news/prismml-releases-ternary-bonsai-2-27b