Ornith-1.5-35B-A3B

Thinking Control & Reasoning Mode — Ornith-1.5-35B-A3B

Ornith-1.5-35B-A3B is a reasoning model: by default the assistant turn opens with a <think>...</think> block before the final answer.

Reasoning Parser

The serving recipes enable a reasoning parser (qwen3) so the chain-of-thought is returned in a separate reasoning_content field:

• vLLM: --reasoning-parser qwen3
• SGLang: --reasoning-parser qwen3

Tool-Call Parser

A tool-call parser surfaces the model's <tool_call> blocks as OpenAI-style tool_calls:

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

Chat Template

The model uses a modified Qwen chat template to ensure consistency between training and inference. The custom template is available at: https://huggingface.co/ornith-ai/Ornith-1.5-35B-A3B/blob/main/chat_template.jinja

Model Ecosystem — Ornith-1.5-35B-A3B

Model Tree

Ornith-1.5-35B-A3B has the following community-derived variants on HuggingFace:

Quantization Variants

Notable official quantization repositories:

• FP8: ornith-ai/Ornith-1.5-35B-A3B-FP8
• GGUF: ornith-ai/Ornith-1.5-35B-A3B-GGUF (BF16 gguf: 71.1 GB, mmproj: 903 MB)
• NVFP4: ornith-ai/Ornith-1.5-35B-A3B-NVFP4

Collection

Part of the Ornith-1.5 collection (12 items, updated 5 days ago, 115 total).

Citation — Ornith-1.5-35B-A3B

@misc{ornith_1_5,
    title = {{Ornith-1.5}: From Self-Scaffolding to Self-Improvement},
    url = {https://ornith.ai/ornith_1_5.html},
    author = {{Ornith Team}},
    year = {2026}
}

Blog post: https://ornith.ai/ornith_1_5.html

Recommended Sampling Parameters & Best Practices — Ornith-1.5-35B-A3B

Sampling Parameters

Key Guidelines

• Reasoning model: By default the assistant turn opens with a <think>...</think> block before the final answer. The serving recipes enable a reasoning parser so the chain-of-thought is returned in a separate reasoning_content field.
• Tool-call parser: Use qwen3_xml (vLLM) or qwen3_coder (SGLang) to surface the model's <tool_call> blocks as OpenAI-style tool_calls.
• Reasoning parser: Use qwen3 for both vLLM and SGLang.
• Context window: 262,144 tokens native. Only enable YaRN RoPE scaling when your workload genuinely needs longer context.
• YaRN factor sizing: Target window ≈ factor × 262,144. For 524K tokens, use factor: 2.0; for 1M tokens, use factor: 4.0.
• GPU memory: 2× 80GB GPUs recommended for 256K context with gpu-memory-utilization 0.90 (vLLM) or mem-fraction-static 0.85 (SGLang).

Context Length & Long-Context Extension — Ornith-1.5-35B-A3B

Native Context

Ornith-1.5-35B-A3B handles context windows of up to 262,144 tokens natively.

Extended Context via YaRN RoPE Scaling

When a task's combined input and output must go beyond the native limit, the effective window can be extended with RoPE scaling — YaRN is the validated technique, already built into both vLLM and SGLang.

With a scaling factor of 4.0, the usable window grows to roughly 1M tokens.

Method 1: Edit checkpoint's config.json

{
    "rope_scaling": {
        "rope_type": "yarn",
        "factor": 4.0,
        "original_max_position_embeddings": 262144
    }
}

Method 2: Override at launch time

vLLM:

VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 vllm serve ornith-ai/Ornith-1.5-35B-A3B ... \
    --hf-overrides '{"rope_scaling": {"rope_type": "yarn", "factor": 4.0, "original_max_position_embeddings": 262144}}' \
    --max-model-len 1000000

SGLang:

SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1 python -m sglang.launch_server ... \
    --json-model-override-args '{"rope_scaling": {"rope_type": "yarn", "factor": 4.0, "original_max_position_embeddings": 262144}}' \
    --context-length 1000000

Note: Open-source runtimes implement YaRN statically: the same scaling factor is applied to every request regardless of its length, which can slightly hurt quality on ordinary-length inputs. Only enable rope_scaling when your workload genuinely needs the longer window, and size factor to match it — the target window is roughly factor × 262,144. For requests topping out around 524,288 tokens, factor: 2.0 is the better setting.

Chat Completions API Usage — Ornith-1.5-35B-A3B

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-1.5-35B-A3B",
    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

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-1.5-35B-A3B",
    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"}

The model emits well-formed function calls that the server parses into the standard tool_calls field. Streaming is also supported. Any OpenAI-compatible SDK (Python, Node.js, etc.) or curl can be pointed at the same /v1/chat/completions endpoint.

Deployment & Serving Recipes — Ornith-1.5-35B-A3B

Ornith-1.5-35B-A3B is a ~35B mixture-of-experts model with ~3B activated parameters per token (≈70 GB in bf16). The recipes below stand up an OpenAI-compatible server on 2× 80GB GPUs to leave headroom for the 256K context.

Runtime Requirements

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

vLLM

vllm serve ornith-ai/Ornith-1.5-35B-A3B \
    --served-model-name Ornith-1.5-35B-A3B \
    --host 0.0.0.0 --port 8000 \
    --tensor-parallel-size 2 \
    --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 ornith-ai/Ornith-1.5-35B-A3B \
    --served-model-name Ornith-1.5-35B-A3B \
    --host 0.0.0.0 --port 8000 \
    --tp 2 \
    --context-length 262144 \
    --mem-fraction-static 0.85 \
    --tool-call-parser qwen3_coder \
    --reasoning-parser qwen3

Ollama

ollama run hf.co/ornith-ai/Ornith-1.5-35B-A3B-GGUF

llama.cpp

llama-server -hf ornith-ai/Ornith-1.5-35B-A3B-GGUF --port 8000 -c 262144

Unsloth Studio

from unsloth import FastLanguageModel
model, tokenizer = FastLanguageModel.from_pretrained(
    "ornith-ai/Ornith-1.5-35B-A3B",
    max_seq_length=262144,
    load_in_4bit=True,
)

Agent Framework Integration

Works with Hermes Agent, OpenClaw, Atomic.chat, and any OpenAI-compatible agent framework. Point the framework at your Ornith server via OPENAI_BASE_URL and OPENAI_API_KEY.

Benchmark Results — Ornith-1.5-35B-A3B

All results are averaged over five independent runs.

Coding

Reasoning

Agentic

Evaluation Notes

• Terminal-Bench 2.1 (Terminus-2): Harbor/Terminus-2 framework, parser=json, temp=1.0, top_p=1.0, 128K context, 4-hour timeout, 32 CPU cores, 48GB RAM, 5 runs averaged.
• Terminal-Bench 2.1 (Claude Code): Claude Code 2.1.126, parser=json, temp=1.0, top_p=1.0, max_new_tokens=131072, 5 runs averaged.
• SWE-Bench Verified/Pro/Multilingual: OpenHands harness, temp=1.0, top_p=0.95, 256K context. Anti-hacking: Git history removed, network disabled.
• DeepSWE: Claude Code harness, temp=1.0, top_p=0.95, 256K context.
• SWE Atlas QnA: mini SWE agent harness, temp=1.0, top_p=0.95, 128K context, 5 runs averaged.
• NL2Repo: temp=1.0, top_p=1.0, 400K context, 48K output. GitHub repos and pip packages blocked.
• HLE: Claude 4.6 Opus as judge model.
• MCP-Atlas: 500-task public subset, thinking mode, 10-min timeout, Claude 4.8 Opus as judge.
• Toolathlon-Verified: Official evaluation service, 128K max tokens.
• ClawEval: Agentic code benchmark, temp=0.6, 256K context.

Ornith-1.5-35B-A3B — Overview & Highlights

Ornith-1.5 is a major step toward building foundation models through end-to-end self-improvement. It extends Ornith-1.0 (developed on top of Qwen3.5 and Gemma4 with additional continued pretraining, mid-training, and post-training) by expanding the self-improvement loop from scaffold and rollout optimization to jointly optimizing task generation, scaffold construction, and solution rollouts.

Rather than relying on a fixed set of human-curated tasks and manually designed harnesses, Ornith-1.5 continuously generates new training tasks, discovers effective strategies for solving them, and improves the policy through reinforcement learning.

Key claims

• Mid-size MoE member of the Ornith-1.5 family: ~35B total params, ~3B activated per token
• Significantly outperforms Qwen 3.6-35B across all coding and agentic benchmarks
• Outperforms dense models Gemma 4-31B and Muse Glimmer-30B by wide margins on agentic coding
• Self-improvement loop: jointly optimizes task generation, scaffold construction, and solution rollouts

Blog: https://ornith.ai/ornith_1_5.html