Skip to main content

Ternary Bonsai 2 27B (PrismML, local server)

Ternary-Bonsai-2-27B is a 27B-parameter model (Apache 2.0) built on Qwen3.8-27B whose weights are quantized to {-1, 0, +1} ("ternary"), retaining ~98% of full-precision benchmark performance at a fraction of the footprint (5.9-8.5GB on disk depending on quant, vs. ~54GB in FP16).

Unlike the other local models in OpenTryOn (qwen3.8, kimi-vl, ...), this one does not load in-process via transformers/diffusers — there is no rotated-weight ternary kernel for those stacks. Instead TernaryBonsaiAdapter is a thin OpenAI-compatible HTTP client, the same shape as an LM Studio or Ollama integration: you start the model's own server, and OpenTryOn just calls it. No opentryon[local] / torch install is needed for this adapter.

Capabilities​

Capabilityternary-bonsai-2-27b
ModalitiesText + image → text
Parameters27.36B, ternary-quantized (~1.76 bits/weight effective)
ContextUp to 262,144 tokens
ThinkingOn by default; per-request budget via thinking_budget_tokens, or server-wide --reasoning-budget
LicenseApache 2.0

Why a separate server​

GGUF variants require the PrismML llama.cpp fork (prism-b10658 or newer) — a stock llama.cpp build silently loads a Q2_0 fallback and produces gibberish, since the rotated weight basis isn't in upstream. MLX weights work with a stock mlx-lm install on Apple Silicon.

VariantRepoDisk
GGUF (PTQ1_0)prism-ml/Ternary-Bonsai-2-27B-gguf5.93 GB
GGUF (PQ2_0, default)same repo7.25 GB
MLX 2-bitprism-ml/Ternary-Bonsai-2-27B-mlx-2bit8.49 GB

Start the server​

Follow PrismML's own instructions (llama.cpp fork, prism-b10658+):

# llama.cpp: OpenAI-compatible HTTP on :8080
./scripts/start_llama_server.sh

# or, on Apple Silicon:
mlx_lm.server --model prism-ml/Ternary-Bonsai-2-27B-mlx-2bit --port 8080

Full instructions: docs.prismml.com/bonsai-2-27b.

Installation (OpenTryOn side)​

# No opentryon[local] needed -- this adapter only uses `openai` + `requests`,
# both already core dependencies.

Configuration​

# BONSAI_BASE_URL=http://127.0.0.1:8080/v1  # default
# BONSAI_API_KEY=...                        # only if your server enforces auth

Quick Start​

from tryon.models.ternary_bonsai import TernaryBonsaiAdapter

adapter = TernaryBonsaiAdapter()  # talks to http://127.0.0.1:8080/v1

result = adapter.understand(
    image="garment.jpg",
    prompt="Describe this outfit: color, pattern, style, fit, and material.",
    thinking_budget_tokens=512,
)
print(result["text"])

If the server isn't running, the adapter raises a RuntimeError pointing at the PrismML docs instead of a raw connection-refused traceback.

CLI​

opentryon understand --model ternary-bonsai-2-27b \
  --image garment.jpg --prompt "Describe this outfit."

opentryon understand --model ternary-bonsai-2-27b \
  --prompt "Explain ternary quantization in two sentences." \
  --thinking-budget-tokens 256

# Point at a different local server
opentryon understand --model ternary-bonsai-2-27b \
  --prompt "hi" --base-url http://127.0.0.1:1234/v1

MCP​

Same registry model appears as an MCP tool (no extra wiring):

  • understand_ternary_bonsai_2_27b — no cloud key needed; requires the local server to be running (see above)

API Reference​

TernaryBonsaiAdapter​

class TernaryBonsaiAdapter:
    def __init__(
        self,
        api_key: Optional[str] = None,   # BONSAI_API_KEY, optional
        model: str = "Ternary-Bonsai-2-27B",
        base_url: Optional[str] = None,  # BONSAI_BASE_URL or http://127.0.0.1:8080/v1
    )

Methods​

  • understand(image=None, prompt=..., thinking_budget_tokens=None, max_tokens=None, temperature=None, model=None)

Return dict keys: text, reasoning, model, usage.

References​