Ornith-1.0

Self-improving open-source models for agentic coding — GGUF quants for Ollama

MIT License Qwen 3.5 architecture 9B Dense + 35B MoE 262K context Tool calling Reasoning (thinking)

Quick Start

Pull any quant directly into Ollama:

# 9B — fits a single 12GB GPU ollama pull maxwell1500/ornith-9b:Q4_K_M # 35B MoE — needs CPU offload on 12GB, fits 24GB+ ollama pull maxwell1500/ornith-35b:Q4_K_M # Run it ollama run maxwell1500/ornith-9b:Q4_K_M

Available Quants

Ornith-1.0-9B (Dense, ~8.95B params)

Quant	Size	Ollama Tag
Q4_K_M	5.6 GB	`maxwell1500/ornith-9b:Q4_K_M`
Q5_K_M	6.5 GB	`maxwell1500/ornith-9b:Q5_K_M`
Q6_K	7.4 GB	`maxwell1500/ornith-9b:Q6_K`
Q8_0	9.5 GB	`maxwell1500/ornith-9b:Q8_0`

Ornith-1.0-35B (MoE, ~34.8B params)

Quant	Size	Ollama Tag
Q4_K_M	21.2 GB	`maxwell1500/ornith-35b:Q4_K_M`
Q5_K_M	24.7 GB	`maxwell1500/ornith-35b:Q5_K_M`
Q6_K	28.5 GB	`maxwell1500/ornith-35b:Q6_K`
Q8_0	36.9 GB	`maxwell1500/ornith-35b:Q8_0`

Recommended Parameters

The model card recommends these sampling settings:

Temperature: 0.6 (use 1.0 to reproduce benchmarks)
Top-p: 0.95
Top-k: 20
Context length: 262,144 (native)

This is a reasoning model — the assistant turn opens with a thinking block before the final answer. Ollama handles this automatically via the embedded chat template.

Benchmarks (9B)

Benchmark	Ornith-9B	Qwen3.5-9B	Qwen3.5-35B	Gemma4-12B
Terminal-Bench 2.1 (Terminus)	43.1	21.3	41.4	21.0
Terminal-Bench 2.1 (Claude Code)	40.6	18.9	38.9	—
SWE-bench Verified	69.4	53.2	70.0	44.2
SWE-bench Pro	42.9	31.3	44.6	27.6
SWE-bench Multilingual	52.0	39.7	60.3	32.5
NL2Repo	27.2	16.2	20.5	10.3
Claw-eval Avg	63.1	53.2	65.4	32.5

Benchmarks (35B MoE)

Benchmark	Ornith-35B	Qwen3.5-35B	Qwen3.6-35B	Qwen3.5-397B
Terminal-Bench 2.1 (Terminus)	64.2	41.4	52.5	53.5
Terminal-Bench 2.1 (Claude Code)	62.8	38.9	49.2	48.6
SWE-bench Verified	75.6	70.0	73.4	76.4
SWE-bench Pro	50.4	44.6	49.5	51.6
SWE-bench Multilingual	69.3	60.3	67.2	69.3
NL2Repo	34.6	20.5	29.4	36.8
Claw-eval Avg	69.8	65.4	68.7	70.7

Credits & Attribution

DeepReinforce Team

deepreinforce-ai

Ornith-1.0 was created and trained by the DeepReinforce Team. All credit for the model architecture, training framework, data curation, and benchmark results belongs entirely to them.

Ornith-1.0 employs a self-improving training framework: it uses RL to learn to generate not only solution rollouts but also the scaffolds that drive those rollouts. By jointly optimizing the scaffold and the resulting solution, the model discovers better search trajectories and generates higher-quality solutions.

The model family includes 9B-Dense, 31B-Dense, 35B-MoE, and 397B-MoE variants, all post-trained on top of Qwen 3.5 and Gemma 4.

HuggingFace 9B Model Card 35B Model Card 9B GGUF 35B GGUF Blog Post

This page distributes pre-quantized GGUF files pulled directly from DeepReinforce's HuggingFace repositories and republished on the Ollama registry under the MIT license for convenience. No modifications were made to the model weights or chat template — these are the original quants from deepreinforce-ai/Ornith-1.0-*-GGUF.

Citation

@misc{ornith_9b, title = {{Ornith-1.0-9B}: Agentic Coding, Open to All}, url = {https://deep-reinforce.com/ornith_1_0.html}, author = {{DeepReinforce Team}}, year = {2026} }

@misc{ornith_35b, title = {{Ornith-1.0-35B}: Agentic Coding, Open to All}, url = {https://deep-reinforce.com/ornith_1_0.html}, author = {{DeepReinforce Team}}, year = {2026} }

Ornith-1.0 is MIT licensed. GGUF quants republished by maxwell1500 with no modifications to original weights.

Details

Readme