Bonsai 27B
Ranked 12 of 39 open-weight language models by recommenders
People recommend it as a 27B-class model that fits in little memory and runs on laptops and phones. The common caveat is that it weakens on long tasks and on vision.
People recommend it as a 27B-class model that fits in little memory and runs on laptops and phones. The common caveat is that it weakens on long tasks and on vision.
141 people gave an opinion
- recommend it
- 89
- criticize it
- 52
- approval share
- 63%
Recommended in 9 of 13 months
- November 2025: not recommended
- December 2025: not recommended
- January 2026: recommended
- February 2026: not recommended
- March 2026: recommended
- April 2026: recommended
- May 2026: recommended
- June 2026: recommended
- July 2026: recommended
- August 2026: recommended
- September 2026: recommended
- October 2026: recommended
- November 2026: not recommended (month in progress)
Why it doesn’t carry the badge
- 89 people recommend itNeeds 10 or more, met.
- 1.7 recommenders for every criticNeeds 2 to 1 or better, not met.
- Recommended in 9 of 13 monthsNeeds 4 or more, met.
The same three rules apply to every product, and the badge can’t be bought. How the badge works
What people say about it
The reasons people give, with the number of comments that give each one. Open a reason to read what they said.
Why people recommend it
Each mark is one comment
Fits in little memory
“My brief experiments with the ternary version suggest that it broadly meets their claim to be a 27B model that fits in much less RAM, that is for sure.”
Comment from 15 Jul 2026 in “Bonsai 27B (1-bit LLM): The First 27B-Class Model to Run on a Phone” “they are genuinely 1-bit (trained from scratch at 1-bit, not quantized down to 4-bit), they actually fit”
Comment from 12 Apr 2026 in “Show HN: Local LLM on a Pi 4 controlling hardware via tool calling” “this new Bonsai is nice because the original was already exciting for making use of low VRAM devices”
Comment from 17 Sep 2026 in “Bonsai 2 27B: Near-Lossless Compression in a 9x Smaller Footprint” - Comment from 27 Oct 2026 in “Survival LLM: AI assistant and emergency library that runs on your own computer”
- Comment from 3 Aug 2026 in “Qwen3.8-Max: A New Bar for Coding and Cowork”
- Comment from 23 Jul 2026 in “Run large language models at home, BitTorrent‑style”
Strong at math and coding
“this model crushes Gemma 4 12b on math and coding, is slightly worse on knowledge and tool calling, and is significantly worse on vision tasks”
Comment from 14 Jul 2026 in “Bonsai 27B (1-bit LLM): The First 27B-Class Model to Run on a Phone” “it is quite an impressive model! I've been grilling it on some deep learning/computer vision stuff and it's aced everything so far.”
Comment from 15 Jul 2026 in “Bonsai 27B (1-bit LLM): The First 27B-Class Model to Run on a Phone” “shockingly smart for a tiny model (but also not Good Enough, by my above definition”
Comment from 15 Apr 2026 in “CPUs Aren't Dead. Gemma2B Out Scored GPT-3.5 Turbo on Test That Made It Famous” - Comment from 1 Apr 2026 in “Show HN: 1-Bit Bonsai, the First Commercially Viable 1-Bit LLMs”
Strong size-to-performance tradeoff
“The interleaved scale factor approach of Bonsai is a much more flexible approach at almost the same cost.”
Comment from 1 Apr 2026 in “Show HN: 1-Bit Bonsai, the First Commercially Viable 1-Bit LLMs” “the inference story on commodity hardware becomes compelling in a way that INT4 quants are not”
Comment from 1 Apr 2026 in “Show HN: 1-Bit Bonsai, the First Commercially Viable 1-Bit LLMs” “quantization and pruning make the model better at the size-performance pareto”
Comment from 15 Jul 2026 in “Bonsai 27B (1-bit LLM): The First 27B-Class Model to Run on a Phone” Good for classification and summarization
“These are not models you'd want to directly interact with. but these models can be very useful for things like classification or simple summarization or translation tasks.”
Comment from 31 Mar 2026 in “Show HN: 1-Bit Bonsai, the First Commercially Viable 1-Bit LLMs” “They’re fairly useful as constrain domain classifiers due to the low memory requirements”
Comment from 18 Sep 2026 in “Bonsai 2 27B: Near-Lossless Compression in a 9x Smaller Footprint”
What people warn about
Each mark is one comment
Claims overstated or cherry-picked
“introducing something like this without thorough benchmarking + model comparison is even more of a red flag to me.”
Comment from 1 Apr 2026 in “Show HN: 1-Bit Bonsai, the First Commercially Viable 1-Bit LLMs” “A bit misleading to say they take 14x less memory, no one is doing inference with 16-bit models.”
Comment from 1 Apr 2026 in “Show HN: 1-Bit Bonsai, the First Commercially Viable 1-Bit LLMs” “The benchmarks they chose are rather cherry picked to not include long context or difficult ones”
Comment from 18 Sep 2026 in “Bonsai 2 27B: Near-Lossless Compression in a 9x Smaller Footprint” Breaks down on long tasks
“The most common failure mode is that it goes into loops and you come back to find it exhausted the output length”
Comment from 18 Sep 2026 in “Bonsai 2 27B: Near-Lossless Compression in a 9x Smaller Footprint” “it's amazing they work as well as they do. Use it for any longer task and they fall apart spectacularly”
Comment from 17 Sep 2026 in “Bonsai 2 27B: Near-Lossless Compression in a 9x Smaller Footprint” Needs special runtime or WebGPU
“The experience is vastly improved but still requires a custom llama.cpp fork to use as of right now.”
Comment from 4 Oct 2026 in “Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s” “The online demos require WebGPU so Firefox on mobilr and privacy enhanced browsers will break.”
Comment from 31 May 2026 in “1-Bit Bonsai Image 4B Image Generation for Local Devices” Weaker on world knowledge and vision
“this model crushes Gemma 4 12b on math and coding, is slightly worse on knowledge and tool calling, and is significantly worse on vision tasks”
Comment from 14 Jul 2026 in “Bonsai 27B (1-bit LLM): The First 27B-Class Model to Run on a Phone” “The V1 was noticeably weak on world knowledge but here the 3.8 base model is geared more towards reasoning than world knowledge anyway”
Comment from 17 Sep 2026 in “Bonsai 2 27B: Near-Lossless Compression in a 9x Smaller Footprint”
Other open-weight language models people recommend
The four most recommended other open-weight language models, with the same counts for each, so you can compare like for like. The tick on each bar marks the 2 to 1 line the badge needs.
| Product | Recommend | Criticize | Approval | Months | Owned | Badge |
|---|---|---|---|---|---|---|
| Qwen 3.8 27B | 1,173 | 426 | 73% | 13 | 2 mo32 owners | Approved |
| Z.ai GLM | 1,043 | 384 | 73% | 13 | 2 mo52 owners | Approved |
| Kimi | 931 | 359 | 72% | 13 | 1 mo26 owners | Approved |
| Gemma 4 | 713 | 249 | 74% | 13 | 1 mo14 owners | Approved |
| Bonsai 27BThis page | 89 | 52 | 63% | 9 | — | — |
What people said
40 of the 179 comments that mention it, newest first. Each links to the original comment.
“You can run bonsai2 27b on basically any gaming GPU from the last 15 years”
“Someone else in this thread recommended Ternary-Bonsai-27B, which I'll also try.”
“The experience is vastly improved but still requires a custom llama.cpp fork to use as of right now.”
“They’re fairly useful as constrain domain classifiers due to the low memory requirements”
“The most common failure mode is that it goes into loops and you come back to find it exhausted the output length”
“try running PTQ2_0 (needs a bit more vram, but is about 2x faster on my 6700 XT)”
“This model might, even without this, be amazing for short lived agents that work via generations / have changing tasks.”
“The benchmarks they chose are rather cherry picked to not include long context or difficult ones”
“this new Bonsai is nice because the original was already exciting for making use of low VRAM devices”
“The V1 was noticeably weak on world knowledge but here the 3.8 base model is geared more towards reasoning than world knowledge anyway”
“for me, it works pretty darn well, with 20-35 tok/s depending on how full that window is”
“it's amazing they work as well as they do. Use it for any longer task and they fall apart spectacularly”
“model fits in 4GB, and with 100k of content model+kv cache fits in 8GB”
“the new Prism Bonsai 27B ternary model is 6GB, runs on any modern Mac”
“Works with the Bonsai-27B-1bit-mlx model on my mac. Edit: ... also it is a fail.”
“quantization and pruning make the model better at the size-performance pareto”
“it is quite an impressive model! I've been grilling it on some deep learning/computer vision stuff and it's aced everything so far.”
“My brief experiments with the ternary version suggest that it broadly meets their claim to be a 27B model that fits in much less RAM, that is for sure.”
“this is truly a paradigm shift - this will kill a whole range of startups in Europe”
“this model crushes Gemma 4 12b on math and coding, is slightly worse on knowledge and tool calling, and is significantly worse on vision tasks”
“I've been waiting for them to start scaling ternary models for over a year”
“It very clearly is not useless and performs a lot better than one might expect.”
“The KV-cache memory usage also seems remarkably frugal, even at the full context length.”
“Maybe bonsai 8b would make the duo, if you do try it, pls post here as I'm a bit curious too.”
“The 1 bit didn't work, as it has a known bug with XCode 24.5 and I wasn't in the mood for installing 24.4 alongside.”
“The immediate load (less than 200ms on my machine through a slow connection) is quite pleasant”
“Yes its a huge deal because these are starting to get bound by memory bandwidth not compute.”
“doesn't it blow your mind that there exists a 1 Gigabyte file/program that can generate any image”
“The online demos require WebGPU so Firefox on mobilr and privacy enhanced browsers will break.”
“The small model is impressive but if you front load a 1.8GB text encoder model, the savings aren't quite as useful.”
“impressive, combines a couple techniques that I always wanted the frontier models to have”
“I have bonsai-8B running with llama-cpp and am attempting to build a harness for it”
“shockingly smart for a tiny model (but also not Good Enough, by my above definition”
“they are genuinely 1-bit (trained from scratch at 1-bit, not quantized down to 4-bit), they actually fit”
“A bit misleading to say they take 14x less memory, no one is doing inference with 16-bit models.”
“the inference story on commodity hardware becomes compelling in a way that INT4 quants are not”
“The interleaved scale factor approach of Bonsai is a much more flexible approach at almost the same cost.”
“I ran my custom agentic SQL debugging benchmark against it and I'm impressed.”
“introducing something like this without thorough benchmarking + model comparison is even more of a red flag to me.”
“These are not models you'd want to directly interact with. but these models can be very useful for things like classification or simple summarization or translation tasks.”
Counts come from 1,795,640 Hacker News comments, November 2025 to November 2026. How we count. Report a mistake.