OsaurusAI/Ornith-1.5-9B-JANG_4D

JANG_4D MLX bundle of ornith-ai/Ornith-1.5-9B — the recommended default — best size/quality balance.

Ornith 1.5 is an agentic coding / reasoning VLM built on a hybrid gated-delta linear attention + full attention backbone (3:1), with a 27-layer vision tower and native video support.

Bundle

Field	Value
Source	ornith-ai/Ornith-1.5-9B
Architecture	qwen3_5 / Qwen3_5ForConditionalGeneration
Size on disk	5.93 GiB
Layers	32
Hidden size	4096
Context	262,144
Shards	2
Bit distribution	{4: 86, 5: 192, 6: 24, 8: 32}

How it was quantized

Three calibration methods, all driven by one capture pass — the per-input-channel second moment E[x_c^2] is simultaneously the Hessian diagonal, the imatrix weighting and the AWQ salient-channel statistic.

Method	What it does here
Hessian-trace allocation	Bits assigned by measured tr(H)·‖W‖²_F per module, not by tensor name. The vision tower scores higher than the text MLP on this model, which a name-based profile gets backwards.
imatrix refit	Activation-weighted affine fit replacing RTN codes — mean weighted rel-err 0.0432.
AWQ	Salient-channel scaling (alpha=0.25), absorbed into the producing RMSNorm across 64 norm groups / 184 projections.

Tensors whose in_features is divisible by no MLX group size (the 27 vision linear_fc2 at 4304) stay fp16.

Modalities

Modality	Status
Text	supported
Vision	supported — 333 vision-tower tensors, preprocessor_config.json + processor_config.json ride with the bundle
Video	supported — video_preprocessor_config.json present; verified end-to-end
Audio	not supported. The tokenizer defines `<

Reasoning

Reasoning is ON by default — the no-kwarg generation prompt is byte-identical to enable_thinking=True and ends <|im_start|>assistant\n<think>\n.

It is toggleable, but note how: enable_thinking=False does not remove the think block, it prefills an empty closed one (<think>\n\n</think>\n\n). A parser testing merely for the presence of a <think> block will find one in both modes — test whether it has content.

There are no reasoning_effort tiers on this model family (unlike Qwen3.8). History <think> blocks are preserved unconditionally. Reasoning parser: qwen3; tool parser: qwen3_coder.

Sampling

Both presets from the vendor card are stamped into jang_config.json, and the coding preset is also written to generation_config.json so the two files agree.

Ornith 1.5 is an agentic coding model (SWE-bench Verified 79, Terminal-Bench 2.1 67.8), so this bundle defaults to the coding preset. Upstream's own generation_config.json ships the general numbers (temp 1.0, presence 1.5) — use sampling_modes.general if you want parity with the vLLM/Transformers defaults.

Preset	temp	top_p	top_k	min_p	presence	repetition
general	1.0	0.95	20	0.0	1.5	1.0
coding (default)	0.6	0.95	20	0.0	0.0	1.0

Stop tokens: [248046, 248044] (<|im_end|>, <|endoftext|>).

Speculative decoding (MTP)

The 9B checkpoint declares mtp_num_hidden_layers: 1 but ships no mtp.* weights, so there is no MTP head to preserve (mtp_mode: metadata_only_missing_weights). Native MTP exists on the 35B-A3B member of this family.

Credits

JANG quantization by Jinho Jang — eric@osaurus.ai

Base model: ornith-ai/Ornith-1.5-9B by Ornith AI.