Ternary Bonsai 2 27B — Uncensored Heretic (NInfer)

A native NInfer artifact (.ninfer) of the Heretic-decensored Ternary Bonsai 2 27B — the uncensored variant, packaged for the consolidated NInfer engine with DFlash2 speculative decoding, MTP, and the Qwen3.8-27B vision tower in a single file.

This fills a gap in the ecosystem: uncensored Bonsai 2 previously existed only in GGUF/MLX form, which no NInfer build can load — and which cannot provide DFlash2 at all. This artifact gives the NInfer audience the de-refused model at full NInfer speed.

Provenance

Upstream lineage: PrismML's ternary Bonsai 2 is derived from Qwen3.8-27B. Apache-2.0 throughout.

Contents

1,193 objects: ternary t2_g128_fp16 language projections, BF16/FP32 GDN controls, Q4/Q5 quantized vision tower (27 blocks + merger), MTP head, the full DFlash2 adapter stack (feature projection, 5 draft layers, conv kernels, candidate-selector codebooks), and the ternary proposal head. 9,520,051,456 bytes — byte-size identical to the official v3 artifact.

Requirements

• The consolidated NInfer line (iamwavecut/ninfer-3090 or a descendant build). The upstream ninfer releases do not decode t2_g128_fp16 weights.
• An NVIDIA GPU with roughly ≥20 GiB free for the configuration below (RTX 3090/4090/5090 class; smaller cards can shrink --kv-capacity / context accordingly).

Example launch

ninfer-serve.exe \
  --model Ternary-Bonsai-2-27B-Heretic-ninfer.ninfer \
  --model-id bonsai2-27b-heretic \
  --spec dflash2 --draft-tokens 7 \
  --kv-dtype rk4v4-e8 --kv-capacity 1048576 --max-context 1048576 \
  --gdn-state-fp16 --max-concurrency 2 --cors \
  --port 11435

Vision variant: add --vision --vision-residency overlay --vision-max-merged 12288.

Measured (RTX 5090, sm_120a build, rk4v4-e8 KV, 1,048,576-token window, thinking off)

All engine-reported (timings) figures; decode rows are sustained multi-thousand-token runs.

Draft acceptance caveat resolved: an early smoke test suggested ~20% acceptance vs ~60% on stock, but that compared different content. Matched-content runs show the ablation costs essentially nothing in throughput — the DFlash2 adapter works fine on the edited weights.

• Smoke-tested, not benchmarked. No perplexity or eval-suite numbers are claimed here.
• Retrieval degrades at the far end of 1M context — inherited from the base model; treat ~500–700K as the dependable window.
• No safety layer. This model is deliberately decensored and will answer requests the official model refuses. Deployment needs its own access controls and policies — nothing in the weights enforces them.

Integrity

SHA256SUMS contains the checksum for the artifact file. Verify before serving:

certutil -hashfile Ternary-Bonsai-2-27B-Heretic-ninfer.ninfer SHA256   # Windows
sha256sum Ternary-Bonsai-2-27B-Heretic-ninfer.ninfer                   # Linux

Credits

• PrismML — Ternary Bonsai 2 base model and ternary format
• OS-Software — Heretic decensoring baked onto the ternary lattice
• Qwen team — Qwen3.8-27B base, vision tower, tokenizer
• WaveCut / iamwavecut — the NInfer ternary engine line and conversion recipe
• ProCreations — MTP head and DFlash2 adapter weights