Ornith-1.0-35B-A3B

Architecture Overview

Ornith-1.0-35B is a Mixture-of-Experts (MoE) model built on the qwen3_5_moe architecture.

Key Specifications

Base Models

Post-trained on top of Gemma 4 and Qwen 3.5 with additional continued pretraining, mid-training, and post-training.

Modalities

• Input: text, image
• Output: text
• The model is multimodal (image-text-to-text).

Languages

• English (en)
• Multi-language support

Model Variants in Family

Context Length

• Native context length: 262,144 tokens (256K)
• Extended context: Not specified (None in DB)

Serving Configuration

• vLLM: --max-model-len 262144
• SGLang: --context-length 262144
• llama.cpp: -c 262144
• Unsloth: max_seq_length=262144

Evaluation Context Usage

Self-Improving Training Framework

Ornith-1.0 employs a novel self-improving training framework using reinforcement learning.

Key Concepts

• Joint Optimization of Scaffold and Rollout: Ornith-1.0 uses RL to learn to generate not only solution rollouts but also the scaffold that drives those rollouts. By jointly optimizing the scaffold and the resulting solution, the model discovers better search trajectories and generates higher-quality solutions.
• Base Models: Post-trained on top of Gemma 4 and Qwen 3.5 with additional continued pretraining, mid-training, and post-training.
• Model Family: Available in 9B-Dense, 31B-Dense, 35B-MoE, and 397B-MoE variants.

Training Stages

1. Continued Pretraining (CPT): Additional pretraining on top of base models.
2. Mid-training: Intermediate training stage between pretraining and post-training.
3. Post-training: Including RL for scaffold and rollout optimization.
4. Scaffold and Rollout Optimization: The model learns to generate both the scaffold (search strategy) and the solution (rollout), jointly optimized via RL.

License

Ornith-1.0-35B is released under the MIT License.

• MIT licensed, globally accessible, and free from regional limitations.
• Commercial use is allowed.
• The model is fully open-source with no regional restrictions.

Reasoning Model Behavior

Ornith-1.0-35B is a reasoning model: by default the assistant turn opens with a … block before the final answer.

Reasoning Content Separation

When using vLLM or SGLang with a reasoning parser (--reasoning-parser qwen3), the chain-of-thought is returned in a separate reasoning_content field:

• reasoning_content: Holds the trace (chain-of-thought)
• content: Holds the final answer

Tool Call Parsing

The serving recipes enable a tool-call parser so the model's <tool_call> blocks are surfaced as OpenAI-style tool_calls:

• vLLM: --tool-call-parser qwen3_xml
• SGLang: --tool-call-parser qwen3_coder

Manual Parsing

When using Transformers directly, split on the `` marker:

text = tokenizer.decode(output_ids, skip_special_tokens=True)
if "" in text:
    reasoning, answer = text.split("", 1)
    reasoning = reasoning.replace("", "").strip()
    answer = answer.strip()
else:
    reasoning, answer = "", text.strip()

Recommended Sampling Parameters

Ornith-1.0-35B is a reasoning model. Recommended sampling parameters for optimal performance:

Evaluation-Specific Settings

• SWE-Bench (OpenHands): temp=1.0, top_p=0.95, 256K context
• Terminal-Bench (Terminus-2): temp=1.0, top_p=1.0, 128K context
• Terminal-Bench (Claude Code): temp=1.0, top_p=1.0, max_new_tokens=131072
• SWE Atlas: temp=1.0, top_p=0.95, 128K context
• NL2Repo: temp=1.0, top_p=1.0, 400K context, 48K output
• ClawEval: temp=0.6, 256K context

Citation

If you find our work helpful, feel free to give us a cite.

@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}
}

Source: Ornith Blog

Agent Framework Integrations

Ornith-1.0-35B exposes an OpenAI-compatible endpoint with tool calling, working out of the box with standard agent frameworks.

Supported Frameworks

• Hermes Agent: Point OPENAI_BASE_URL at your Ornith server, set MODEL=deepreinforce-ai/Ornith-1.0-35B.
• Atomic.chat / Ollama / llama.cpp: Load a GGUF build of Ornith. llama-server -hf deepreinforce-ai/Ornith-1.0-35B-GGUF --port 8000 -c 262144 or ollama run hf.co/deepreinforce-ai/Ornith-1.0-35B-GGUF.
• OpenClaw: Set OPENAI_MODEL=deepreinforce-ai/Ornith-1.0-35B with your Ornith server endpoint.
• Unsloth Studio: pip install unsloth, load with FastLanguageModel.from_pretrained("deepreinforce-ai/Ornith-1.0-35B", max_seq_length=262144, load_in_4bit=True).
• OpenHands: pip install openhands-ai, set LLM_MODEL=openai/deepreinforce-ai/Ornith-1.0-35B, route through LiteLLM.
• OpenCode: Register Ornith as a provider in ~/.config/opencode/opencode.json with @ai-sdk/openai-compatible.

MCP Server Example

import os
from openai import OpenAI

client = OpenAI(
    base_url=os.getenv("OPENAI_BASE_URL", "http://localhost:8000/v1"),
    api_key=os.getenv("OPENAI_API_KEY", "EMPTY"),
)
tools = [{
    "type": "function",
    "function": {
        "name": "run_shell",
        "description": "Run a shell command and return its output.",
        "parameters": {"type": "object", "properties": {"command": {"type": "string"}}, "required": ["command"]},
    },
}]
response = client.chat.completions.create(
    model="deepreinforce-ai/Ornith-1.0-35B",
    messages=[{"role": "user", "content": "List the Python files in the current directory."}],
    tools=tools, temperature=0.6, top_p=0.95,
)

Model Tree for ornith-ai/Ornith-1.0-35B

Ornith-1.0-35B serves as a base model for a wide community ecosystem:

Quantizations are available for multiple runtimes: llama.cpp, LM Studio, Jan, and Ollama.

The model is part of the Ornith-1.0 Collection — a family of open-source LLMs specialized for agentic coding (8 items total).

Local Inference with Hugging Face Transformers

For a quick local test or offline generation, load Ornith-1.0-35B directly with Transformers (requires transformers >= 5.8.1).

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "deepreinforce-ai/Ornith-1.0-35B"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, dtype="auto", device_map="auto")

messages = [{"role": "user", "content": "Write a Python function is_prime(n). Keep it short."}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
generated = model.generate(
    **inputs, max_new_tokens=512, do_sample=True,
    temperature=0.6, top_p=0.95, top_k=20,
)
output_ids = generated[0][inputs.input_ids.shape[1]:]
content = tokenizer.decode(output_ids, skip_special_tokens=True)
print(content)

Splitting Reasoning from Answer

The reply contains a … reasoning block followed by the final answer. Parse on the `` marker:

text = tokenizer.decode(output_ids, skip_special_tokens=True)
if "" in text:
    reasoning, answer = text.split("", 1)
    reasoning = reasoning.replace("", "").strip()
    answer = answer.strip()
else:
    reasoning, answer = "", text.strip()

Using Ornith-1.0-35B via the Chat Completions API

Once a vLLM or SGLang server is running, talk to it with any OpenAI-compatible client.

Basic Usage

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")

response = client.chat.completions.create(
    model="Ornith-1.0-35B",
    messages=[{"role": "user", "content": "Write a one-line Python lambda that squares a number."}],
    temperature=0.6, top_p=0.95, max_tokens=1024,
)

message = response.choices[0].message
print("reasoning:", getattr(message, "reasoning_content", None))
print("answer:", message.content)

Tool Calling

Ornith-1.0-35B emits well-formed function calls that the server parses into the standard tool_calls field.

tools = [{
    "type": "function",
    "function": {
        "name": "get_weather",
        "description": "Get the current weather for a city",
        "parameters": {
            "type": "object",
            "properties": {"city": {"type": "string"}},
            "required": ["city"],
        },
    },
}]

response = client.chat.completions.create(
    model="Ornith-1.0-35B",
    messages=[{"role": "user", "content": "What is the weather in Paris right now?"}],
    tools=tools, tool_choice="auto",
    temperature=0.6, max_tokens=2048,
)

tool_call = response.choices[0].message.tool_calls[0]
print(tool_call.function.name, tool_call.function.arguments)
## -> get_weather {"city": "Paris"}

Any OpenAI-compatible SDK (Python, Node.js) or curl can target the same /v1/chat/completions endpoint.

Serving Ornith-1.0-35B

Ornith-1.0-35B requires recent runtimes:

• Transformers ≥ 5.8.1
• vLLM ≥ 0.19.1
• SGLang ≥ 0.5.9

The recipes below stand up an OpenAI-compatible server on a single 8×80GB GPU node (tensor-parallel 8). Adjust --tensor-parallel-size / --tp to the number of GPUs available.

vLLM

vllm serve deepreinforce-ai/Ornith-1.0-35B \
    --served-model-name Ornith-1.0-35B \
    --tensor-parallel-size 8 \
    --host 0.0.0.0 --port 8000 \
    --max-model-len 262144 \
    --gpu-memory-utilization 0.90 \
    --enable-prefix-caching \
    --enable-auto-tool-choice --tool-call-parser qwen3_xml \
    --reasoning-parser qwen3 \
    --trust-remote-code

SGLang

python -m sglang.launch_server \
    --model-path deepreinforce-ai/Ornith-1.0-35B \
    --served-model-name Ornith-1.0-35B \
    --tp 8 \
    --host 0.0.0.0 --port 8000 \
    --context-length 262144 \
    --mem-fraction-static 0.85 \
    --tool-call-parser qwen3_coder \
    --reasoning-parser qwen3

Hugging Face Transformers

Requires transformers >= 5.8.1. See the Transformers installation guide.

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "deepreinforce-ai/Ornith-1.0-35B"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, dtype="auto", device_map="auto")

Agentic Coding Benchmark Results

Evaluation Methodology

• Terminal-Bench 2.1 (Terminus-2): Harbor/Terminus-2 framework, parser=json, temperature=1.0, top_p=1.0, 128K context, 4h timeout, 32 CPU cores, 48GB RAM, avg of 5 runs.
• Terminal-Bench 2.1 (Claude Code): Claude Code 2.1.126, parser=json, temperature=1.0, top_p=1.0, max_new_tokens=131072, avg of 5 runs.
• SWE-Bench Verified/Pro/Multilingual: OpenHands harness, temp=1.0, top_p=0.95, 256k context window.
• SWE Atlas QnA/RF/TW: mini SWE agent harness, temp=1.0, top_p=0.95, 128K context, avg of 5 runs.
• NL2Repo: temperature=1.0, top_p=1.0, 400K context, 48K output, anti-hacking filters.
• ClawEval: Agentic code benchmark over real-user task distributions, temp=0.6, 256K context.

Ornith-1.0 Key Highlights

Ornith-1.0 is a self-improving family of open-source models for agentic coding.

• State-of-the-Art Coding Agents: Available in 9B-Dense, 31B-Dense, 35B-MoE, and 397B-MoE (post-trained on top of Gemma 4 and Qwen 3.5), achieving state-of-the-art performance among open-source models of comparable size on coding benchmarks such as Terminal-Bench 2.1, SWE-Bench, NL2Repo and OpenClaw.
• Self-Improving Training Framework: Ornith-1.0 employs RL to learn to generate not only solution rollouts, but also the scaffold that drive those rollouts. By jointly optimizing the scaffold and the resulting solution, the model discovers better search trajectories and generates higher-quality solutions.
• Licence: MIT licensed, globally accessible, and free from regional limitations.

The 35B-MoE variant is the lightweight member of the Ornith family, designed for efficient single-GPU deployment.