Blackfrost-AI/TERNARY-BONSAI-2-27B-DERISKED-GGUF

Verified creator Blackfrost-AI verified
🤗 Hugging Face sourceimage-text-to-textapache-2.063 GBGGUFHF checksums availableupdated today
No torrent yet

TERNARY BONSAI 2 27B — 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
└── prism-ml/Ternary-Bonsai-2-27B-gguf
    ├── TERNARY BONSAI 2 27B — DERISKED BF16 GGUF
    └── TERNARY BONSAI 2 27B — 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 113–123 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.