by Prism MlReleased 18 September 2026
Ternary Bonsai 2 27B reads text and pictures and writes back text, and it keeps about 260 pages in mind at once. It is PrismML's compressed version of Qwen3.8 27B: the files shrink to roughly 8.5 GB while it keeps 98.2% of the original's average score on PrismML's 14 thinking tests. Use it for coding, maths or questions about a picture on modest hardware; it thinks before it answers, at the highest effort.
You pay for what each request uses. No subscription, no minimum.
One message, the ordinary use
Every rate
| Meter | ↑ you send | ↓ comes back |
|---|---|---|
| Text tokens | 24,96 DA | 166,42 DA |
| Cached tokens read | 12,48 DA |
per 1M tokens
One company runs this model. OpenDunes sends every request to it. DA per 1M tokens: ↑ what you send · ↓ what comes back.
You pay the price of the provider that served your request. Prices read 15:48 · speed over the last 30 minutes · uptime over 3 days.
Real requests, real replies. Try any of them yourself — one click, pre-filled.
What's the total on this invoice, and when is it due?
invoice-scan.jpg · photoThe total is 42 300,00 DA, due 5 September 2026. Worth a double-check: the handwritten day could read as 8 instead of 3.
Try in Chat LabTurn this customer email into JSON with intent, urgency and order_id.
customer-email.txtTry your own request — same model, pre-selected.
Open Chat LabSwap the base URL and the key. Your OpenAI code runs unchanged.
import os
from openai import OpenAI
client = OpenAI(
base_url="https://opendunes.com/v1",
api_key=os.environ["OPENDUNES_API_KEY"],
)
resp = client.chat.completions.create(
model="prism-ml/ternary-bonsai-2-27b",
messages=[{"role": "user", "content": "Two lines to welcome a shop in Oran."}],
)
print(resp.choices[0].message.content)Runs this request once on your balance.
Standard OpenAI fields, grouped by what they control. Anything not listed is accepted and ignored.
Shaping the reply
How long, how varied, where to stop.
max_tokenstemperaturetop_ptop_kstopseedfrequency_penaltypresence_penaltyrepetition_penaltylogprobstop_logprobsReasoning
Let it think before answering — and see the thinking.
reasoningreasoning_effortinclude_reasoningStructured output
Get JSON you can parse instead of prose.
response_formatstructured_outputsTools
Let the model call your functions.
toolstool_choiceAbout 0,11 DA per message. ↑ 24,96 DA · ↓ 166,42 DA per 1M tokens.
No. You pay per use from your balance, with no subscription. Try it in Chat at the same price.
Ternary Bonsai 2 27B reads text and pictures and writes back text, and it keeps about 260 pages in mind at once. It is PrismML's compressed version of Qwen3.8 27B: the files shrink to roughly 8.5 GB while it keeps 98.2% of the original's average score on PrismML's 14 thinking tests. Use it for coding, maths or questions about a picture on modest hardware; it thinks before it answers, at the highest effort.
Ternary Bonsai 2 9B costs 0,062 DA less for the same message.
Deposit dinars with your Edahabia or CIB card. Every model, Ternary Bonsai 2 27B included, is paid from that balance. No foreign card, no conversion.
Prism Ml names no languages for Ternary Bonsai 2 27B. Arabic, French and Darja: not stated.
It reads text and images. Up to 262K tokens in one request, about 260 pages. It writes up to 25 000 words at once.
Yes, its weights are public. Open weights on Hugging Face
Prism Ml. Released 18 September 2026.
This model accepts text and pictures as input, processing up to 260 pages of information and generating responses up to 65 pages. It is best suited for complex tasks that require careful reasoning and the use of external tools. A key feature is its open nature and highly efficient internal design. However, it is limited by its inability to search the web or access external live data.
From what its maker and two benchmark sites publish, and where it stands on OpenDunes. Read 7 Nov 2026.
Thinks first
Works through the problem before it answers.
Uses your tools
Can call functions your code provides.
Fills your JSON
Answers in the JSON shape you give it.
Searches the web No.
Can look things up online while it answers.
Cheaper on repeats
A prompt it saw a moment ago costs less the second time.
Open weights
The model's weights are public.
Run a private coding assistant on one GPU
Thinks first
Solve a maths problem step by step
Thinks first
Call your tools from an on-device agent
Uses your tools