Bonsai-2-27B-1.75bit-JANG

JANG-affine bundle of prism-ml/Ternary-Bonsai-2-27B-mlx-2bit, Prism ML's ternary Bonsai 2 27B (built on Qwen3.8-27B). This is the dense-packed edition: the same exact ternary weights as Bonsai-2-27B-Ternary-JANG, stored at Prism's PTQ1_0 density (1.75 bits/weight: 5 trits per byte plus one fp16 scale per 128-weight group, no stored bias) and expanded losslessly into MLX's native 2-bit affine kernels at load time. Output is bit-identical to the ternary bundle; only the download shrinks. Proper affine JANG storage, not JANGTQ, MXTQ, or a codebook sidecar format.

OsaurusAI · osaurus.ai · JANG source

Bundle

Property	Value
Architecture	Dense Qwen3.8-27B conditional-generation VLM (64 blocks: 48 GatedDeltaNet + 16 full attention)
JANG profile	JANG_AFFINE_TERNARY_PACKED
Text matrices	ternary {−s, 0, +s}, 26 bytes of packed trits + fp16 scale per 128-group on disk (1.75 bpw), 2-bit slots in memory, exact
Weight basis	blockwise Hadamard rotation (block 1024, explicit signs), applied to activations at runtime
Vision linears	6-bit affine, group size 128
Norms and recurrent-state tensors	float32 passthrough (as in the source)
Weight shards	6.08 GiB
Context	262,144 tokens
Modalities	text, image, video
Audio	not supported

Embeddings, the untied language-model head, full-attention projections, GatedDeltaNet projections, and MLP matrices are ternary. Bonsai is dense: it has no routed experts or router tensors. The ternary codes decode to exactly the source {−scale, 0, +scale} groups.

The bundle contains the original tokenizer, tokenizer config, the Qwen3.8 chat template (thinking, reasoning_effort, tools), image and video processor configs, the Prism hadamard.json sidecar, source license and notice. EOS metadata is normalized to <|im_end|> (248046).

Runtime

The language model is stored in a Hadamard-rotated basis. Stock mlx_lm / mlx_vlm loaders return wrong output silently because they skip the activation transform. Use Osaurus or a vMLX build with JANG Hadamard and ternary-packed support (osaurus.json names the minimum Osaurus version); the loader expands the packed trits (codes and scales unchanged, biases materialized as −scale), applies the activation transform from the bundle's declared contract, and refuses to load if any module or sign vector is missing. Runtime memory and speed equal the ternary bundle.

vmlx serve OsaurusAI/Bonsai-2-27B-1.75bit-JANG --host 127.0.0.1 --port 8000

OpenAI-compatible chat requests support text, image_url, and video_url content parts, tool definitions, and chat_template_kwargs for enable_thinking and reasoning_effort (low, medium, xhigh; default xhigh).

Sampling defaults follow the Qwen3.8 card that Prism also recommends: thinking mode temperature 1.0, top_p 0.95, top_k 20; instruct mode temperature 0.7, top_p 0.80, top_k 20, presence_penalty 1.5.

Verification

Verified on 2026-09-17 through the vMLX Python server on an Apple M5 Max with 128 GB unified memory.

Gate	Result
Logit parity vs the ternary bundle	PASS — all 402 expanded modules bit-identical, logits bit-identical on 3 prompts
Logit parity vs Prism's reference loader (via the ternary bundle)	PASS — argmax agreement 1.0 at every position on 4 prompts, identical greedy continuations
Single-turn text, thinking off	PASS — Paris
Thinking on (reasoning_effort=medium)	PASS — closed think block, correct 391
Multi-turn	PASS — exact ORCHID-4729 recall and combination
Long context	PASS — buried fact recalled from a 10,655-token prompt
Image	PASS — red background with centered blue square; green circle plus exact OCR of overlaid text
Video	PASS — red frames followed by blue frames
Tool calling	PASS — get_weather call emitted and tool result folded into the final answer

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

Quantization notes

• 402 language-model modules (embedding, 64 layers, untied head) are the source ternary codes and scales, packed 5 trits per byte (26 bytes per 128-group) without re-quantization; biases are always −scale and are not stored. There is no full-precision source for these weights, so AWQ, imatrix and GPTQ do not apply.
• 83 eligible vision linears use native 6-bit affine storage; blocks.N.mlp.linear_fc2 (input 4304) and the patch/position embeddings stay float16.
• 699 norms, GatedDeltaNet state projections, convolutions, biases and incompatible vision tensors pass through in their source precision.
• No tq_packed, tq_norms, mxtq_bits, or jangtq_runtime.safetensors artifacts are present.

License and attribution

Apache-2.0. See LICENSE and NOTICE.txt. This repository is a repacked conversion of the linked Prism ML Bonsai 2 checkpoint; the ternary weights are Prism ML's work.