Instructions to use Blackfrost-AI/TERNARY-BONSAI-2-27B-DERISKED-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- llama.cpp
How to use Blackfrost-AI/TERNARY-BONSAI-2-27B-DERISKED-GGUF with llama.cpp:
Install (macOS, Linux)
curl -LsSf https://llama.app/install.sh | sh # Start a local OpenAI-compatible server with a web UI: llama serve -hf Blackfrost-AI/TERNARY-BONSAI-2-27B-DERISKED-GGUF:BF16 # Run inference directly in the terminal: llama cli -hf Blackfrost-AI/TERNARY-BONSAI-2-27B-DERISKED-GGUF:BF16
Install from WinGet (Windows)
winget install llama.cpp # Start a local OpenAI-compatible server with a web UI: llama serve -hf Blackfrost-AI/TERNARY-BONSAI-2-27B-DERISKED-GGUF:BF16 # Run inference directly in the terminal: llama cli -hf Blackfrost-AI/TERNARY-BONSAI-2-27B-DERISKED-GGUF:BF16
Use pre-built binary
# Download pre-built binary from: # https://github.com/ggerganov/llama.cpp/releases # Start a local OpenAI-compatible server with a web UI: ./llama-server -hf Blackfrost-AI/TERNARY-BONSAI-2-27B-DERISKED-GGUF:BF16 # Run inference directly in the terminal: ./llama-cli -hf Blackfrost-AI/TERNARY-BONSAI-2-27B-DERISKED-GGUF:BF16
Build from source code
git clone https://github.com/ggerganov/llama.cpp.git cd llama.cpp cmake -B build cmake --build build -j --target llama-server llama-cli # Start a local OpenAI-compatible server with a web UI: ./build/bin/llama-server -hf Blackfrost-AI/TERNARY-BONSAI-2-27B-DERISKED-GGUF:BF16 # Run inference directly in the terminal: ./build/bin/llama-cli -hf Blackfrost-AI/TERNARY-BONSAI-2-27B-DERISKED-GGUF:BF16
Use Docker
docker model run hf.co/Blackfrost-AI/TERNARY-BONSAI-2-27B-DERISKED-GGUF:BF16
- LM Studio
- Jan
- vLLM
How to use Blackfrost-AI/TERNARY-BONSAI-2-27B-DERISKED-GGUF with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "Blackfrost-AI/TERNARY-BONSAI-2-27B-DERISKED-GGUF" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "Blackfrost-AI/TERNARY-BONSAI-2-27B-DERISKED-GGUF", "messages": [ { "role": "user", "content": [ { "type": "text", "text": "Describe this image in one sentence." }, { "type": "image_url", "image_url": { "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg" } } ] } ] }'Use Docker
docker model run hf.co/Blackfrost-AI/TERNARY-BONSAI-2-27B-DERISKED-GGUF:BF16
- Ollama
How to use Blackfrost-AI/TERNARY-BONSAI-2-27B-DERISKED-GGUF with Ollama:
ollama run hf.co/Blackfrost-AI/TERNARY-BONSAI-2-27B-DERISKED-GGUF:BF16
- Unsloth Desktop
- Pi
How to use Blackfrost-AI/TERNARY-BONSAI-2-27B-DERISKED-GGUF with Pi:
Start the llama.cpp server
# Install llama.cpp: brew install llama.cpp # Start a local OpenAI-compatible server: llama serve -hf Blackfrost-AI/TERNARY-BONSAI-2-27B-DERISKED-GGUF:BF16
Configure the model in Pi
# Install Pi: npm install -g @earendil-works/pi-coding-agent # Add to ~/.pi/agent/models.json: { "providers": { "llama-cpp": { "baseUrl": "http://localhost:8080/v1", "api": "openai-completions", "apiKey": "none", "models": [ { "id": "Blackfrost-AI/TERNARY-BONSAI-2-27B-DERISKED-GGUF:BF16" } ] } } }Run Pi
# Start Pi in your project directory: pi
- Docker Model Runner
How to use Blackfrost-AI/TERNARY-BONSAI-2-27B-DERISKED-GGUF with Docker Model Runner:
docker model run hf.co/Blackfrost-AI/TERNARY-BONSAI-2-27B-DERISKED-GGUF:BF16
- Lemonade
How to use Blackfrost-AI/TERNARY-BONSAI-2-27B-DERISKED-GGUF with Lemonade:
Pull the model
# Download Lemonade from https://lemonade-server.ai/ lemonade pull Blackfrost-AI/TERNARY-BONSAI-2-27B-DERISKED-GGUF:BF16
Run and chat with the model
lemonade run user.TERNARY-BONSAI-2-27B-DERISKED-GGUF-BF16
List all available models
lemonade list
- Hermes Agent
How to use Blackfrost-AI/TERNARY-BONSAI-2-27B-DERISKED-GGUF with Hermes Agent:
Start the llama.cpp server
# Install llama.cpp: brew install llama.cpp # Start a local OpenAI-compatible server: llama serve -hf Blackfrost-AI/TERNARY-BONSAI-2-27B-DERISKED-GGUF:BF16
Configure Hermes
# Install Hermes: curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash hermes setup # Point Hermes at the local server: hermes config set model.provider custom hermes config set model.base_url http://127.0.0.1:8080/v1 hermes config set model.default Blackfrost-AI/TERNARY-BONSAI-2-27B-DERISKED-GGUF:BF16
Run Hermes
hermes
- Atomic Chat
- OpenClaw
How to use Blackfrost-AI/TERNARY-BONSAI-2-27B-DERISKED-GGUF with OpenClaw:
Start the llama.cpp server
# Install llama.cpp: brew install llama.cpp # Start a local OpenAI-compatible server: llama serve -hf Blackfrost-AI/TERNARY-BONSAI-2-27B-DERISKED-GGUF:BF16
Configure OpenClaw
# Install OpenClaw: npm install -g openclaw@latest # Register the local server and set it as the default model: openclaw onboard --non-interactive --mode local \ --auth-choice custom-api-key \ --custom-base-url http://127.0.0.1:8080/v1 \ --custom-model-id "Blackfrost-AI/TERNARY-BONSAI-2-27B-DERISKED-GGUF:BF16" \ --custom-provider-id llama-cpp \ --custom-compatibility openai \ --custom-text-input \ --accept-risk \ --skip-health
Run OpenClaw
openclaw agent --local --agent main --message "Hello from Hugging Face"
TERNARY BONSAI 2 27B โ DERISKED BF16 + PQ2_0 GGUF
Two behaviorally modified multimodal GGUF operating points: reference BF16 and compact native PQ2_0.
Release status
| Item | Status |
|---|---|
| BF16 language model | Included; operator coherence review accepted |
| PQ2_0 language model | Included; operator coherence review accepted |
| Multimodal projectors | BF16 and Q8_0 included |
| Refusal evaluation | 0/450 |
| Comprehensive derivative benchmark | Pending |
| Deployment kit | Included for one NVIDIA RTX PRO 6000 Blackwell GPU |
| Modification recipe | Proprietary and intentionally not distributed |
DERISKED identifies the release family. It is not a claim of zero refusals, guaranteed safety, or suitability for unsupervised deployment. Refusal measurements will be added only after the pending evaluation is complete.
Why this release exists
This repository places the accepted high-precision and compact variants together so users can choose between reference-quality BF16 and the substantially smaller PQ2_0 artifact without losing the model's optional image-input path. Both artifacts share the same architectural lineage but have distinct checksums and must be treated as separate inference targets.
Included artifacts
| Artifact | Purpose | Bytes | Approximate size | SHA-256 |
|---|---|---|---|---|
TERNARY-BONSAI-2-27B-DERISKED-BF16.gguf |
High-precision language model | 53,808,408,928 | 50.11 GiB | 16e8a683e0c83747911dc2e6dbe9da759e5f9af56d59e8b6889c0287b16e2a7b |
TERNARY-BONSAI-2-27B-DERISKED-PQ2_0.gguf |
Compact native ternary language model | 7,206,168,928 | 6.71 GiB | 32eb8f0ddfb8714d7ea9d10903c6dbe56c3b508f876c7280c6072dafcb1db7d5 |
TERNARY-BONSAI-2-27B-MMPROJ-BF16.gguf |
BF16 multimodal projector | 931,145,856 | 0.87 GiB | e287342d92332fa3577ed1d42e921dac9370c08da58ba9337fa450f6cc76cfd7 |
TERNARY-BONSAI-2-27B-MMPROJ-Q8_0.gguf |
Compact multimodal projector | 629,246,976 | 0.59 GiB | 6807ede61d570bb86ba34b756a0fa109edc33668604de867c6ea6d8f1d631903 |
Use the BF16 projector with the BF16 language model and the Q8_0 projector with PQ2_0 unless independently validating another pairing.
Model specifications
| Property | Value |
|---|---|
| Architecture | Qwen3.8-derived hybrid-attention causal language model |
| Parameters | Approximately 27.36B total; 64 transformer blocks, per the parent card |
| BF16 representation | High-precision GGUF language weights |
| PQ2_0 representation | Native ternary g128 language weights with FP16 group scales |
| Modalities | Text; image input when the corresponding projector is loaded |
| Upstream context specification | 262,144 tokens |
| Validated release context | 32,768 tokens |
| Validated concurrency | One parallel slot / one generation stream |
| License | Apache-2.0, inherited from the immediate parent |
The upstream context specification is not a claim that these derivatives were validated at the full 262K window.
Lineage
Qwen/Qwen3.8-27B
โโโ prism-ml/Ternary-Bonsai-2-27B-gguf
โโโ TERNARY BONSAI 2 27B โ DERISKED BF16 GGUF
โโโ TERNARY BONSAI 2 27B โ DERISKED PQ2_0 GGUF
| Stage | Provenance |
|---|---|
| Foundation | Qwen/Qwen3.8-27B, as identified by the immediate parent's metadata |
| Immediate parent | prism-ml/Ternary-Bonsai-2-27B-gguf |
| Parent revision | 6ed5e12bf84b7a63069882c91dd9e9218647d17b |
| Parent BF16 artifact | Ternary-Bonsai-2-27B-F16.gguf, SHA-256 f6f3b2c9b41956c34b379ec7c301dc936bc38d79b3c24c83388dd7d76000c180 |
| Parent PQ2_0 artifact | Ternary-Bonsai-2-27B-PQ2_0.gguf, SHA-256 3907dc1658db1f78a9826bf8d5bcb8dc65db0d466388937af57f2294fae62ec1 |
| Derivative stage | Weight-level behavioral modification; proprietary process not distributed |
| Runtime lineage | Prism ML llama.cpp fork, pinned in the deployment kit |
Tokenizer, processor, GGUF metadata, architecture, and projector lineage follow the immediate parent except where the released checksums identify derivative language-model artifacts.
Artifact verification
sha256sum *.gguf
Compare all four results against the Included artifacts table before serving. A matching filename without a matching checksum is not the validated release artifact.
Behavioral and capability evaluation
- Both selected operating points passed operator/HITL coherence review.
- The selected PQ2_0 artifact is the currently accepted non-looping compact candidate.
- The BF16 artifact produced approximately 27.05 decode tokens/second in the documented single-stream smoke configuration.
- PQ2_0 operational observations were approximately 113โ123 decode tokens/second in the documented single-stream configuration.
- Refusal evaluation is pending; no refusal-rate claim is made in this release.
- Comprehensive coding, security, long-context, and multimodal derivative results are not yet published.
- Scores reported for the upstream model do not automatically transfer to these modified artifacts.
These observations are not controlled cross-model benchmarks and are not performance guarantees.
Prompting, templates, and tools
The validated launcher enables the GGUF Jinja chat template and does not inject a deployment-specific system prompt. Use role-structured messages and allow enough generation budget for reasoning-heavy work.
The runtime exposes an OpenAI-compatible chat endpoint. It does not provide an application tool executor: the client remains responsible for schema handling, tool execution, result reinjection, timeouts, and conversation state.
Deployment
The tested single-GPU path is included in DEPLOYMENT_KIT_RTX_PRO_6000_BLACKWELL/. The launcher supports both released language-model/projector pairs and preserves the exact validated serving flags.
Limitations and security responsibility
- PQ2_0 is an aggressive low-bit representation and can differ from BF16.
- Published validation covers 32,768 tokens and one stream, not the full upstream context claim.
- Image input requires the corresponding projector and a compatible Prism runtime.
- Refusal evaluation is pending.
- Treat generated text, code, URLs, tool arguments, file paths, and commands as untrusted until independently validated.
- The model is not an authorization, policy-enforcement, or security boundary.
- Deployers are responsible for authentication, rate limiting, network isolation, audit logging, data handling, and compliance.
License and disclaimer
Released under Apache-2.0, subject to the included license, notices, and immediate parent's terms. Review all upstream terms before redistribution or deployment.
This research artifact is provided without warranties. Users are responsible for validating accuracy, security, suitability, and legal compliance for their use case.
Use repository Discussions for reproducible artifact reports. Never post credentials, personal information, or private infrastructure details.
- Downloads last month
- 2,653
2-bit
8-bit
16-bit