ornith-ai/Ornith-1.0-9B — an RL post-trained agentic coding model on a Qwen3.5-9B backbone — converted to the OpenVINO™ IR format with weights compressed to INT4 by NNCF.

Quantization recipe

optimum-cli export openvino \
  --model ornith-ai/Ornith-1.0-9B \
  --task image-text-to-text \
  --weight-format int4 --sym --group-size 128 --ratio 1.0 \
  Ornith-1.0-9B-int4-symg128-ov

Symmetric g128, ratio 1.0, data-free (no AWQ, no scale-estimation).
Data-free is not a preference here, it is the only option: the VLM export path resolves calibration sets by name from PREDEFINED_VISUAL_LM_DATASETS (a raw corpus raises TypeError: unhashable type: 'list'), contextual is the sole accepted name, and its image bucket now returns HTTP 403 AllAccessDisabled. So AWQ / scale-estimation are unreachable for this architecture at present.
A four-recipe ablation on the sibling 1.5 release found symmetric beat asymmetric, and g128 beat g64 by a wide margin — coarser grouping was better, counter to the usual intuition.
ACTIVATIONS_SCALE_FACTOR left at the baked 8.0; verified coherent on long generations (hidden size 4096 does not need the bump larger Gemma/Ministral do).

Requires OpenVINO GenAI 2026.3.0.0 or newer (qwen3_5 support). The IR is VLM-shaped, so it loads through VLMPipeline even for text-only use.

Usage

python

import openvino_genai as ov_genai

pipe = ov_genai.VLMPipeline("Ornith-1.0-9B-int4-symg128-ov", "GPU")
cfg = ov_genai.GenerationConfig()
cfg.max_new_tokens = 512
cfg.do_sample = True
cfg.temperature = 0.6      # vendor-recommended for precise coding
cfg.top_p = 0.95
cfg.top_k = 20
print(pipe.generate("Write a Python function that merges overlapping intervals.",
                    generation_config=cfg))

It is a reasoning model: output opens with a thinking trace and emits a closing </think> (the chat template pre-fills the opener), so split on that tag before showing the answer. Tool calls use the hermes <tool_call> wrapper with a Qwen3-Coder XML body (<function=name><parameter=k>v</parameter></function>), not hermes JSON — parsers expecting JSON inside the wrapper will see zero calls.

Measured behaviour

Scored on a per-task-type benchmark (each cell pass/fail; code is executed against hidden assertions) on an Intel Core Ultra 155H Arc iGPU, OpenVINO GenAI 2026.3.0.0, at the vendor-recommended coding operating point:

suite	score	notes
codegen	11/12
agent-loop	7/7	full multi-turn tool discipline, incl. deep chains
analysis	4/4
edit (single-shot)	0/2	see below
autocomplete-fim	n/a	skipped for VLM-shaped IRs

Known behavioural limit. On single-shot edit tasks it calls read_file instead of editing — even when the file contents are already in the prompt and a system prompt instructs it to edit directly. This held at every decoding setting tested (greedy, 0.2, 0.6, thinking on). Its Qwen3.5-9B base passes the same cell 2/2, so this is a trade made by the agentic post-training: multi-turn loop discipline gained, single-shot compliance lost. Use it in an agent loop that lets it read, edit and test across turns — which is what it scores 7/7 on.

Lowering the sampling temperature to improve that cell is a net loss: it buys the edit cell but costs agent-loop and analysis (4/4 → 2/4, at 7× the wall-clock), because the reasoner loops at low temperature.

HarmenWessels/Ornith-1.0-9B-int4-symg128-ov

Ornith-1.0-9B — OpenVINO int4 (symmetric, group size 128)

Quantization recipe

Usage

Measured behaviour

HarmenWessels/
Ornith-1.0-9B-int4-symg128-ov