Modelpublic

OsaurusAI/Bonsai-27b-Ternary-JANG

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

2likes398downloads

Model Card

Bonsai-27b-Ternary-JANG

Ternary 2-bit-slot JANG-affine conversion of prism-ml/Ternary-Bonsai-27B-unpacked. This is proper affine JANG storage, not JANGTQ, MXTQ, or a codebook sidecar format.

OsaurusAI · osaurus.ai · JANG source

Bundle

Property	Value
Architecture	Dense Qwen3.5 conditional-generation VLM, 27B
JANG profile	`JANG_AFFINE_TERNARY_2BIT`
Text matrices	ternary 2-bit slots, group size 128
Vision linears	4-bit affine, group size 64
Norms and state tensors	float16 passthrough
Weight shards	7.477 GiB
Modalities	text, image, video
Audio	not supported

Embeddings, the untied language-model head, full-attention projections, GatedDeltaNet projections, and MLP matrices use the same discrete text profile. Bonsai is dense: it has no routed experts or router tensors. Ternary codes represent the source {-scale, 0, +scale} groups directly with affine scales and biases.

The bundle contains the original tokenizer vocabulary, tokenizer config, Qwen chat template with thinking and tool support, image processor config, and video processor config. EOS metadata is normalized to <|im_end|> (248046). Source license and notice files are included.

Runtime

Use a vMLX Python build containing schema-2 discrete JANG-affine storage and mixed-precision vision support. Stock mlx_lm does not implement one-bit affine storage and does not honor this bundle's per-module mixed precision.

bash

vmlx serve OsaurusAI/Bonsai-27b-Ternary-JANG --host 127.0.0.1 --port 8000

OpenAI-compatible chat requests support text, image_url, and video_url content parts. The included chat template supports enable_thinking and tool definitions.

Verification

Verified on 2026-07-14 with vMLX Python on an Apple M5 Max with 128 GB unified memory.

Gate	Result
Single-turn text	PASS — coherent Paris answer
Multi-turn	PASS — exact `ORCHID-4729` recall through three turns
Hybrid cache	PASS — paged KV plus SSM companion cache hit
Image	PASS — identified red background and centered blue square
Video	PASS — identified red frames followed by blue frames

The conversion report is included as jang_affine_report.json; authoritative per-tensor storage metadata is in jang_config.json.

Quantization notes

—498 discrete text matrices use the selected binary/ternary affine profile.
—83 eligible vision linears use native 4-bit affine storage.
—603 norms, state tensors, convolutions, biases, and incompatible vision tensors remain float16.
—No tq_packed, tq_norms, mxtq_bits, or jangtq_runtime.safetensors artifacts are present.

License and attribution

Apache-2.0. See LICENSE, LICENSE.txt, and NOTICE.txt. This repository is a quantized conversion of the linked PrismML Bonsai source checkpoint.