The zoo’s first agentic-coding / self-scaffolding model. Source:
deepreinforce-ai/Ornith-1.0-9B
(MIT, released 2026-06-25) — DeepReinforce’s Ornith 1.0 family trains the model to
jointly solve coding tasks AND construct the orchestration scaffold that guides the
solution; the 9B dense is its phone-/Mac-class member (the 397B MoE sibling matches
frontier assistants on SWE-Bench Verified). An image-text-to-text checkpoint; this
card is the text decoder.
Architecturally it is a Qwen3.5 hybrid decoder (model_type qwen3_5) — the same
family as the shipped 0.8B / 2B / Qwen3.6-27B — so it rides the
pipelined engine with zero new export code
(the stock script + --hf-id):
full_attention_interval 4 → 8 full / 24 linear);MLP(12288), hidden 4096, untied 248320-vocab lm_head (at the
checkpoint root — the loader’s untied branch, like the 27B);model.visual.*, 333 tensors) and an MTP
config stub but no MTP weights; the text-decoder load skips vision cleanly.⬇️ Converted .aimodel bundle: ornith_1_0_9b_decode_int8hu_block32_sym/
(9.8 GB, full LanguageBundle incl. tokenizer + chat template; decode-only loop-free,
max_ctx 8192). HF upload user-gated.
| check | result |
|---|---|
| oracle (fp32 HF eager, margin-validated prompt) | sweep min top-2 margin 0.205 (most 2–8) |
| fp16 eager teacher-forced (baseline) | 24/24 exact (16 sweep + 8 decode) |
| int8hu eager (ship recipe) | 24/24 exact — zero flips, even at the 0.205 / 0.058-margin knife edges |
engine path (release llm-runner, greedy, raw prompt ids) |
12/12 token-exact ≡ fp32 oracle (int8hu AND int4lin) |
Mac bench int8hu (M4 Max, p128 g256 n3, COREAI_CHUNK_THRESHOLD=1) |
decode 48.3 / prefill 48.5 tok/s (trials ±0.1) |
| Mac bench int4lin (same rig) | decode 58.9 / prefill 59.0 tok/s (+22%, 7.5 GB) |
Ship recipe = the family ship shape: linear int8 per-block-32 symmetric_with_clipping
body (embed/conv/norms fp16) + absmax symmetric per-block-32 int8 on the untied
big-vocab head (clipping crushes outlier head rows — see the knowledge page). Effective
per-token read ≈ 7.7 GB → the 48 tok/s is bandwidth-bound and healthy.
Chat path: the bundled Qwen chat template works as-is (llm-runner --prompt wraps it;
fluent assistant-style code analysis).
int4lin (linear per-block-32, head fp16) gates 24/24 exact on the same eager
teacher-forced rig — including the 0.205 / 0.058-margin knife edges — and the engine
bundle reproduces the oracle greedy 12/12. That is a first for the Qwen3.5 family:
0.8B failed int4 at every scheme, the 2B failed linear per-block-32, the 27B was
borderline (one confident flip). Quantization sensitivity is a model property; this
RL-trained 9B tolerates 4-bit where its siblings don’t. Measured: 7.5 GB bundle,
58.9 tok/s decode on M4 Max (+22% over int8hu) — this shape is in the
gemma4/LFM “+~24%” class, not the qwen-2B flat class.
Why int8hu is still the default ship: every instrument we ran is clean for int4, but they are all short-context (48 gated tokens). The gemma4-VL port showed int4 clipping error can compound invisibly until real context lengths (272-token gate caught flips a 27-token gate missed). Until a long-context gate says otherwise, int8hu carries the quality claim; int4lin is one flag away and fully verified at short range.
int8 at 9.8 GB exceeds the entitled jetsam ceiling (~6.4 GB on iPhone 17 Pro) — Mac
ship for v1. int4lin (7.5 GB measured) helps but not enough as-is. The arithmetic
route to the phone is int4 body + int8 head (≈ 6.5 GB, at the jetsam edge) or
additionally an in-graph int8 embed table (≈ 5.5 GB, real headroom) — both future
work, the latter needs the gemma4---tbl-style int8 embed gather.
cd coreai-models && .venv/bin/python \
../coreai-models-community/conversion/export_qwen3_5_decode_pipelined.py \
int8hu --head-sym --hf-id deepreinforce-ai/Ornith-1.0-9B --max-ctx 8192
# gate (eager, fp32-oracle margin rule):
# _smoke/gen_ornith9b_ref.py (oracle) + _smoke/test_ornith9b_eager_gate.py
COREAI_CHUNK_THRESHOLD=1 llm-benchmark \
--model exports/ornith_1_0_9b_decode_int8hu_block32_sym -p 128 -g 256 -n 3
Requires the qwen3_5 model overlay + the pipelined-engine extra-states patch
(apps/coreai-pipelined-extra-states.patch) — identical requirements to the shipped
qwen3.5 family; nothing new.
Model weights MIT (upstream deepreinforce-ai/Ornith-1.0-9B). This conversion keeps
the upstream license; see the HF card for the Ornith 1.0 technical report.