Official OsaurusAI JANG_4M build of deepreinforce-ai/Ornith-1.0-9B (MIT) — a vision-language model on a Qwen3.5 hybrid backbone. Mixed-precision affine (JANG_4M); runs on Apple Silicon via Osaurus.

~6.3 GB (from ~18.8 GB bf16) bundle.
JANG_4M: 8-bit affine attention + 4-bit affine MLP (mixed precision, group-size 64); embeddings/head per-module; vision tower kept fp16.
Vision-language (image + text → text).

Architecture


Family	`qwen3_5` (dense, hybrid)
Text layers	32 — 24 Gated-DeltaNet (linear-attention) + 8 full-attention
MoE / dims	hidden 4096 · untied lm_head
Vision	ViT tower (`model.visual`) preserved fp16
Cache	hybrid (GDN state + KV for attention layers)
Parsers	reasoning `qwen3` · tools `qwen`

Running it

JANG bundles use the Qwen3.5 RMSNorm +1 scale_shift applied at runtime, so load them in Osaurus (or the vMLX runtime), which handles it automatically:

osaurus run OsaurusAI/Ornith-1.0-9B-JANG_4M

Note: a plain mlx_lm.generate will not be coherent on a JANG bundle — it omits the +1 norm shift. Use the Osaurus / vMLX runtime (or vmlx_engine.loaders.load_jang), which applies it.

Provenance

Quantization: Osaurus · JANG_4M (8-bit affine attention + 4-bit affine MLP, group-size 64; vision tower fp16) · eric@osaurus.ai