PrismML Ternary Bonsai 27B

27B-class reasoning in ternary weights retaining 95% of full-precision quality

Try now

read docs

About model

Ternary Bonsai 27B is PrismML's low-bit build of Qwen3.6 27B, representing every language weight as one of three values with group-wise FP16 scaling for a true 1.71 bits per weight, roughly 9.4x smaller than the FP16 baseline. The compression runs end-to-end across embeddings, attention projections, MLP projections, and the LM head, with a compact 4-bit vision tower, and the model retains 95% of full-precision quality with an 80.49 average across 15 thinking-mode benchmarks. Math stays within two points of full precision at a 93.40 category average, and the 262K-token context rides on Qwen3.6's hybrid-attention backbone. Available free on Together AI.

‍

FP16 Quality Retained

95%

80.49 average across 15 thinking-mode benchmarks

True Bits per Weight

1.71

End-to-end ternary weights, roughly 9.4x smaller than FP16

Context Window

262K

Hybrid-attention backbone for long-document and repository work

Model key capabilities

Quality-Retaining Ternary Weights: End-to-end ternary language weights at a true 1.71 bits per weight, retaining 95% of the FP16 baseline across 15 thinking-mode benchmarks
Intact Math & Coding: 99.20 MATH-500, 90.84 AIME 2025, and 93.90 HumanEval+, with the math category within two points of full precision
Multimodal & Agentic: Vision input for image and document understanding plus 74.41 BFCL v3 tool calling in the sub-4-bit regime
Production-Ready Infrastructure: 99.9% SLA, available free on Together AI serverless infrastructure

Performance benchmarks

Model	AIME 2025	GPQA Diamond	HLE	LiveCodeBench	MATH500	SWE-bench verified
PrismML Ternary Bonsai 27B	90.84			82.75	99.20
Related open-source models
Competitor closed-source models
Claude Opus 4.6		90.5%	34.2%			78.7%
OpenAI o3		83.3%	24.9%		99.2%	62.3%
OpenAI o1		76.8%			96.4%	48.9%
GPT-4o		49.2%	2.7%	32.3%	89.3%	31.0%

API usage

cURL
Python
Typescript

Endpoint:

Prism-ML/Ternary-Bonsai-27B

curl -X POST https://api.together.xyz/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $TOGETHER_API_KEY" \
  -d '{
    "model": "Prism-ML/Ternary-Bonsai-27B",
    "messages": [{
      "role": "user",
      "content": "Given two binary strings `a` and `b`, return their sum as a binary string"
    }]
  }'

from together import Together

client = Together()
response = client.chat.completions.create(
  model="Prism-ML/Ternary-Bonsai-27B",
  messages=[
  	{
	    "role": "user", 
      "content": "Given two binary strings `a` and `b`, return their sum as a binary string"
    }
 ],
)

print(response.choices[0].message.content)

import Together from "together-ai";

const together = new Together();

async function main() {
  const response = await together.chat.completions.create({
    model: "Prism-ML/Ternary-Bonsai-27B",
    messages: [{
      role: "user",
      content: "Given two binary strings `a` and `b`, return their sum as a binary string"
    }]
  });
  
  console.log(response.choices[0]?.message?.content);
}

main();

Model card
Architecture Overview:
• Derived from Qwen3.6 27B, a hybrid-attention causal language model (~75% linear, ~25% full attention) with the architecture unchanged
• ~27.32B ternary language weights plus a ~461M vision tower in 4-bit NF4
• Ternary g128 format: three-valued weights with one FP16 scale per group of 128, a true 1.71 bits per weight
• 262K-token context window
• Ships with MTP and DSpark speculative-decoding drafter layers trained against the Bonsai 27B target, with lossless verification that preserves the target distribution exactly

Training Methodology:
• Built from the Qwen3.6 27B base with PrismML's low-bit representation applied end-to-end across the language stack
• All variants evaluated under identical infrastructure, decoding, and scoring in thinking mode across 15 benchmarks in six skill categories

Performance Characteristics:
• Thinking-mode average of 80.49 across 15 benchmarks, 95% of the FP16 baseline's 85.07
• Math: 99.20 MATH-500, 96.06 GSM8K, 90.84 AIME 2025
• Coding: 93.90 HumanEval+, 81.22 MBPP+, 82.75 LiveCodeBench
• Agentic and instruction following: 74.41 BFCL v3, 73.61 Tau2-Bench, 85.03 IFEval
• Vision: 68.96 MMMU-Pro, 61.42 OCR Bench v2
• PrismML reports a higher thinking-mode average than a conventional 2-bit Q2_XXS build (73.39) at well under two-thirds of its footprint
‍
Prompting
Together AI API Access:
• Access Ternary Bonsai 27B via Together AI APIs using the endpoint Prism-ML/Ternary-Bonsai-27B
• Authenticate using your Together AI API key in request headers
• Recommended sampling: temperature 0.7, top_p 0.95, top_k 20, the settings behind all reported thinking-mode results
• Supports text and image input with a 262K-token context window, plus JSON mode for structured outputs
• Available free on Together AI serverless infrastructure
‍
Applications & use cases
Cost-Efficient Reasoning Agents:
• Serve 27B-class reasoning and tool calling through the Together AI serverless endpoint at no cost
• Run agent loops that lean on the model's math and coding strength within two points of full precision
• Scale batches and concurrent sessions with headroom from the small memory footprint

Long-Context Document & Code Analysis:
• Load long documents and large codebases into the 262K context for sustained analysis
• Summarize, cross-reference, and reason over material that depends on holding a large working set in context
• Pair long-context reads with structured follow-up queries in the same session

Multimodal Understanding:
• Parse documents, screenshots, and charts through the vision tower alongside text prompts
• Combine OCR-style extraction with downstream reasoning in a single call
• Ground text workflows in supplied images through the Together API
‍

Related models

No items found.

Model specifications

Model data

Model provider
PrismML
Type
Chat
Reasoning
Vision
Features
JSON Mode
Deployment
Serverless
Endpoint
Prism-ML/Ternary-Bonsai-27B
Parameters
27B
Context length
262K
Input price
0.00 / 1M tokens
Input modalities
Text
Image
Output modalities
Text

Released
July 14, 2026
Category
Code

Run in Playground

Quickstart docs

Deploy model

PrismML Ternary Bonsai 27B

About model

API usage

Model card

Prompting

Applications & use cases