# Ternary Bonsai

14 Hacker News commenters recommended Ternary Bonsai and 8 criticized it, November 2025 to November 2026. It was recommended in 3 of 13 months.

**Verdict:** Compact 1.58-bit model people praise for accuracy per byte and fast on-device use. Critics say it hallucinates often and is weak at coding and logic.

## Why people recommend it

- Strong accuracy for its size (3 comments): "Ternary-Bonsai-1.7B achieved 65.1% from 462 MiB" ([source](https://news.ycombinator.com/item?id=47843500)) "Essentially seems to be a very good quantization of that model" ([source](https://news.ycombinator.com/item?id=47845931)) "good-per-byte-but-not-in-absolute-terms quantization" ([source](https://news.ycombinator.com/item?id=47846071))
- Runs well for on-device AI (3 comments): "This is currently the best model for on-device AI and has been used in our three apps." ([source](https://news.ycombinator.com/item?id=47797575)) "it's great progress toward local AI, which I believe will be common in the near future" ([source](https://news.ycombinator.com/item?id=48094276)) "Vram usage was consistently at ~7GiB." ([source](https://news.ycombinator.com/item?id=47844055))
- Fast tokens and first response (2 comments): "it seems to get higher TPS than most other similarly sized models on my hardware" ([source](https://news.ycombinator.com/item?id=48913511)) "time to first token is incredibly low" ([source](https://news.ycombinator.com/item?id=47845998))

## What people warn about

- Weak at coding tasks (2 comments): "works great for basic conversational tasks and not good for one shot coding tasks" ([source](https://news.ycombinator.com/item?id=47845676)) "bonsai is basically trash on real agentic coding work" ([source](https://news.ycombinator.com/item?id=50099500))
- Benchmarks look cherry-picked (2 comments): "Independent testing of prismml suggest quite a capability drop off outside of their cherry picked benchmarks" ([source](https://news.ycombinator.com/item?id=49079975)) "the comparisons read as a bit disingenuous with the other models at 16 bit weights" ([source](https://news.ycombinator.com/item?id=47846469))
- Hallucinates answers often (2 comments): "much like the bonsai tenary one it hallucinates knowledge quite aggressively" ([source](https://news.ycombinator.com/item?id=49176679)) "hallucinates in pretty much every answer" ([source](https://news.ycombinator.com/item?id=47845234))
- Weak logic and literal answers (2 comments): "Fails the car wash test (basic logic seems to be weak in general)" ([source](https://news.ycombinator.com/item?id=47845556)) "This model tends to be annoyingly literal." ([source](https://news.ycombinator.com/item?id=47843959))

## Alternatives hackers recommend

- [Microsoft BitNet](https://hackerapproved.com/p/bitnet.md): 16 recommend
- [Maple-Preview](https://hackerapproved.com/p/maple-preview.md): 6 recommend
- [BitNet b1.58 2B4T](https://hackerapproved.com/p/bitnet-b1-58-2b4t.md): 5 recommend

## Cite as

Hacker Approved, "Ternary Bonsai", https://hackerapproved.com/p/ternary-bonsai/ (counts from 1,795,640 Hacker News comments, November 2025 to November 2026).