Ternary Bonsai 2 27B
3357°
Run this model
Offers / Ads
PrismML | 262,144 token context | $0.07/M input, $0.50/M output. Bonsai 2 27B is a 27B-parameter reasoning model from PrismML derived from Qwen3.8-27B. It supports coding, mathematics, tool calling, and image understanding with a 262K-token context window. Ternary compression shrinks the language-model weights to roughly 8.5 GB while retaining 98.2% of the base model's average score across PrismML's 14 thinking-mode benchmarks, enabling efficient inference on consumer hardware. The model thinks by default and defaults to xhigh reasoning effort.
Guides for this model

Ollama vs LM Studio (2026): which local AI app to use, and which top models each one runs
What is a GGUF file? Quantized models explained (Q4, Q8, fp8, NF4)
How much VRAM do I actually need for local AI?
Captain's notes
Notes from 2026-10-13
Ternary Bonsai 2 27B is a 27B-parameter reasoning model from PrismML, built on a Qwen3.8-27B backbone and compressed to ternary weights so the language model fits in roughly 6 GB. It retains 98.2% of the base model's benchmark average across 14 thinking-mode tests, covering coding, math, tool calling, and vision. If you want a full-class reasoning model running locally on a laptop or a single consumer GPU, this is the most aggressive compression that still holds its shape.
Known for
- ~6 GB language model footprint on a 27B-class reasoner, via true ternary weights
- 98.2% of FP16 benchmark average retained across 14 thinking-mode evaluations
- Runs at ~47 tok/s on an Apple M5 Max laptop via llama.cpp or MLX
- 262K-token context kept practical on-device by a hybrid-attention backbone
- Thinks by default at xhigh effort; math, coding, and tool calling all benchmarked
Watch out for
- Requires a custom llama.cpp fork with ternary kernels; standard builds will not work correctly.
- Vision tower ships as a separate Q8_0 mmproj pack, not bundled in the main GGUF.
- Ternary compression is novel; check PrismML's known-issues page before deploying in production.
Prompts to try
Trace a tricky bug
Here is a Python function that is supposed to return the top-k most frequent words in a list of strings, but it sometimes returns fewer than k results. Walk through the logic step by step, identify the bug, and give me a corrected version with a brief explanation of what was wrong.
```python
from collections import Counter
def top_k_words(texts, k):
words = ' '.join(texts).split()
counts = Counter(words)
return [w for w, _ in counts.most_common() if counts[w] >= k]
```
Solve a logic puzzle
Five houses in a row are each painted a different colour. The owners each drink a different beverage, smoke a different brand, and keep a different pet. Use these clues to determine who keeps the fish: the Brit lives in the red house; the Swede keeps dogs; the Dane drinks tea; the green house is immediately left of the white house; the green house owner drinks coffee; the person who smokes Pall Mall keeps birds; the owner of the yellow house smokes Dunhill; the man in the centre house drinks milk; the Norwegian lives in the first house; the man who smokes Blends lives next to the cat owner; the man who keeps horses lives next to the Dunhill smoker; the owner who smokes BlueMaster drinks beer; the German smokes Prince; the Norwegian lives next to the blue house; the man who smokes Blends ha
Plan a tool-calling agent
You have three tools: search(query) returns web snippets, calculator(expr) evaluates a math expression, and file_write(name, content) saves a file. A user asks: find the current population of Tokyo, divide it by the area of Tokyo in km², round to two decimal places, and save the result as tokyo_density.txt. Write out the exact sequence of tool calls you would make, with arguments, and explain your reasoning at each step.
Compare with
- Claude Opus 5.5API-only model1,000,000 token context · $4.00/M input, $20.00/M output
- Gemini 3.8 FlashAPI-only model1,048,576 token context · $0.83/M input, $4.12/M output
- Claude Sonnet 5.5API-only model1,000,000 token context · $2.00/M input, $10.00/M output
- GPT-5.6 SolAPI-only model1,050,000 token context · $4.00/M input, $20.00/M output
More text models from prism-ml
Questions people ask
- Can I run Ternary Bonsai 2 27B locally, and are the weights open?
- Yes, the weights are open and available as GGUF files on Hugging Face under prism-ml/Ternary-Bonsai-2-27B-gguf. You need PrismML's custom llama.cpp fork for the ternary kernels; standard llama.cpp builds are not supported.
- What does it cost, and how long is the context window?
- Via PrismML's API it is $0.07 per million input tokens and $0.50 per million output tokens, with a 262,144-token context window.
- How does ternary compression compare to a standard Q4 quantisation?
- PrismML reports 84.78 average across 14 thinking-mode benchmarks, well above the conventional IQ2_XXS build at 72.59, and within 0.4 points of a Q4_K_XL build that is roughly three times larger on disk.
Captain's verdictThe most credible case yet that you can run a serious reasoning model on a laptop without meaningful capability loss.
Model card on Hugging Face
Captain writes these notes from the model's own description and model card. Prices and context sizes are the provider's at the time of writing and can change.
Ranking badge
Updated Oct 22, 2026
Made this model? Show this badge on your page.
Paste it into your Hugging Face card or Civitai description.
How this score adds up
This model has no popularity data from any source yet, so it carries a stand-in number below 7,000° that keeps the lists complete. It never ranks above a model with data. The breakdown appears here as soon as a source reports it.
Glossary: the terms used here
- Heat score
- One number for how much a model is used and wanted right now, from 0° to 9,999°. Every list on this site that says hottest is sorted by it and nothing else.
- Floor
- The 7,000° every model with real data starts from. The scale runs from 0° to 9,999°; the part below the floor is reserved for stand-ins (models without data), so that a model with data always ranks above one without. The floor is the same for everyone and decides nothing; the degrees above it are the measurement.
- Download-equivalents
- Everything the score treats as a download, added up: downloads themselves, each torrent seeder as 1,000, each AI Horde image and each run on a generation site as one.
- Log scale
- Counting by tens, not by ones: going from 1,000 to 10,000 downloads adds the same degrees as going from 10,000 to 100,000. It keeps giants from crowding out everything else.
- Likes
- Thumbs up and likes from every source, worth half a download each on the log scale.
- Adoption rate
- How many downloads a day a model gets right now, from the difference between two readings of a source. A quarter of the weight of the download term.
- Measured and estimated
- Measured means two readings of a source exist and the difference is real. Estimated means only today's total is known and the past is spread over the model's life; the site says which one it is showing.
- Seeder
- Someone who keeps a copy of the weights online for others to download over a torrent, or a hosted mirror doing the same. Each one counts as 1,000 downloads.
- AI Horde
- A volunteer network that generates images for free on donated computers. The number of images it has made with a model is public, and each counts as one download.
- Generation sites
- Sites that host their own copies of a model and let people generate with it, such as PixAI, SeaArt and Yodayo. Their downloads count like Civitai's and each generation run as one download.
- OpenRouter tokens
- A text model served through an API, such as Claude or GPT, has no weights to download, so its usage is measured in tokens: the pieces of text it read and wrote through OpenRouter, a gateway many apps use, over the last 30 days. Every 10,000,000 tokens count as one download. Source: OpenRouter (openrouter.ai/rankings), CC BY 4.0.
- Full model (checkpoint)
- The weights themselves, the file that generates on its own. It gets a bonus because add-ons depend on it.
- LoRA and embedding
- Small add-on files that steer someone else's weights toward a style, a character or a concept. They cannot run alone.
- Family and weights
- The base a model is built on, such as SD 1.5, SDXL, Flux or Wan. Every model inherits the release date of its family's weights.
- Age
- Degrees taken off for how long ago the family's weights came out: 150° a year, up to 750°. It keeps old downloads that had years to pile up from freezing the top of the list.
- Family boost
- A small, bounded nudge for LoRAs of a family this site is pushing forward right now, always listed in the breakdown.
- Stand-in score
- A stable number below the floor given to a model no source reports on yet, so lists stay complete. It never ranks above a model with data.
- New! (grace period)
- A model less than 7 days old shows New! instead of a score and holds no place in the rankings. Its score is computed from day one, but a model one day old grows faster than anything, so without the wait the week's #1 would always be whatever came out yesterday rather than what held up for a week.
- Trend
- How a model's growth over a period compares with the typical model's growth over the same period. Faster than typical reads as a rise, slower as a fall.
- Typical model
- The middle of the catalog: half the models with a figure grew faster over the period, half slower.
