Occamy-35B
Evolving Data, Expanding Agent Capability
Introduction
We present Occamy-35B, an efficient agentic model designed for long-horizon decision-making, coordinated multi-tool use, and complex business workflows. Post-trained from Qwen3.6-35B-A3B, Occamy demonstrates strong stability and task-completion capability on challenging Claw-style tasks.
As LLMs evolve from passive responders into agents capable of executing long-horizon tasks in interactive environments, static training data can no longer keep pace with their rapidly changing capability boundaries.
The key innovation behind Occamy is Data RSI (Data Recursive Self-Improvement), a model-driven data evolution framework. For each rollout, the system dynamically composes a new executable task from:
Persona × Tools × Fixtures × Skill x Constraints × Grader. Rollout outcomes are fed back into the data generator: failures reveal capability gaps, while successes expose reusable behavioral patterns. These signals guide subsequent task generation toward the model's evolving capability frontier, forming a closed loop of policy-adaptive data evolution.
Across agentic coding, Claw-style workflows, and real-world e-commerce scenarios, Occamy-35B demonstrates strong and robust capabilities, matching or even surpassing substantially larger models such as GLM-5.2 and DeepSeek-V4 Pro.
Open Source
We will open-source Occamy-35B, together with a subset of our data derived from real-world e-commerce scenarios, to support and advance open research on agentic models.
| Resource | Link |
|---|---|
| Model | occamy-ai/occamy-1.0 |
| Dataset | occamy-ai/occamy-data-1.0 |
Model Performance
We evaluate Occamy-35B across agentic coding, Claw-style workflows, multi-turn tool use, and real-world e-commerce scenarios. It matches or outperforms several substantially larger open-weight and proprietary models, highlighting its strong generalization across both business workflows and agentic environments.
| Benchmark | Occamy-35B-A3B | Qwen3.6-35B-A3B | Gemma 4 31B | Qwen3.6-27B | KAT-Coder-V2.5-Dev | Muse-Glimmer-30B | Claude Opus 4.6 | GLM-5.2 | GPT-5.6 Sol | DeepSeek-V4 Pro | Claude Opus 4.8 | GPT-5.4 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Agent | ||||||||||||
| Claw-Eval Avg | 84.3 | 68.7 | 48.5 | 72.4 | — | — | 80.4 | 81.6 | — | 77.2 | — | 78.4 |
| Claw-Eval Pass³ | 71.3 | 50.0 | 25.0 | 60.6 | — | — | 70.4 | 68.3 | — | 59.8 | — | 60.3 |
| WildClawBench | 50.1 | 40.4 | 37.6 | 43.2 | — | 47.6 | 51.6 | 54.2 | 67.2 | 43.7 | 64.7 | 50.3 |
| τ³-Bench Pass¹ | 31.2 | 11.9 | — | — | — | — | 27.3 | 29.6 | 46.9 | — | 39.7 | 39.4 |
| RealReplicaBench | 37.3 | 19.6 | — | — | — | — | 43.9 | 39.3 | 49.5 | 33.6 | 51.4 | 42.1 |
| Coding & SWE | ||||||||||||
| Terminal-Bench 2.1 | 59.0 | 49.5 | — | — | 41.0 | 51.7 | — | 82.7 | 88.8 | 67.9 | 78.9 | — |
| SWE-bench Pro | 43.5 | 40.4 | — | 53.5 | 46.0 | 51.2 | — | 62.1 | 64.6 | 55.4 | 69.2 | 57.5 |
* Indicates that the model was evaluated by our team.
Terminal-Bench 2.1: Harbor/Terminus-2 harness; 3-hour timeout; 32 CPU/48 GB RAM; temperature=1.0, top_p=0.95, top_k=20, max_tokens=80K, and 256K context; average of two runs.
SWE-bench Pro: Internal agent scaffold with bash and str_replace_edit tools; temperature=1.0, top_p=0.95, and 200K context. We use the unmodified public set. All models were re-evaluated in-house under this identical setup; no officially reported numbers are adopted in this table.
τ³-Bench: We report Banking Knowledge results following the official example configuration, using AllTools retrieval and GPT-5.2 with reasoning_effort=low as the user simulator.
Usage
For production deployment, we recommend serving Occamy-35B with SGLang or vLLM through an OpenAI-compatible API.
SGLang Deployment
SGLang >=0.5.10 is recommended.
Install SGLang:
uv pip install "sglang[all]>=0.5.10"
Launch the server with the Qwen3 reasoning parser and Qwen3-Coder tool-call parser:
python -m sglang.launch_server \
--model-path <ORG>/<MODEL_NAME> \
--host 0.0.0.0 \
--port 8000 \
--tp-size 8 \
--mem-fraction-static 0.8 \
--context-length 262144 \
--reasoning-parser qwen3 \
--tool-call-parser qwen3_coder
Once the server is running, the OpenAI-compatible API is available at:
http://localhost:8000/v1
vLLM Deployment
vLLM >=0.19.0 is recommended.
Install vLLM:
uv pip install "vllm>=0.19.0" --torch-backend=auto
Launch the server with automatic tool selection enabled:
vllm serve <ORG>/<MODEL_NAME> \
--host 0.0.0.0 \
--port 8000 \
--tensor-parallel-size 8 \
--max-model-len 262144 \
--reasoning-parser qwen3 \
--enable-auto-tool-choice \
--tool-call-parser qwen3_coder
Once the server is running, the OpenAI-compatible API is available at:
http://localhost:8000/v1
OpenAI-Compatible API
SGLang and vLLM expose the same OpenAI-compatible API, allowing them to share the same client implementation.
Install the OpenAI Python SDK:
pip install -U openai
Text Generation
from openai import OpenAI
client = OpenAI(
base_url="http://localhost:8000/v1",
api_key="EMPTY",
)
response = client.chat.completions.create(
model="<ORG>/<MODEL_NAME>",
messages=[
{
"role": "user",
"content": "Analyze the current task and provide a complete solution.",
}
],
temperature=1.0,
top_p=0.95,
max_tokens=32768,
extra_body={
"top_k": 20,
},
)
message = response.choices[0].message
print("Thinking:", getattr(message, "reasoning_content", None))
print("Answer:", message.content)
Occamy-35B is primarily designed to operate in Thinking mode. Disabling reasoning through options such as
enable_thinking=Falseor the/nothinkcommand is not recommended.
Function Calling
When the tool-call parser is enabled during deployment, tool definitions can be passed directly through the OpenAI-compatible API:
from openai import OpenAI
client = OpenAI(
base_url="http://localhost:8000/v1",
api_key="EMPTY",
)
tools = [
{
"type": "function",
"function": {
"name": "get_order",
"description": "Retrieve order information using an order ID.",
"parameters": {
"type": "object",
"properties": {
"order_id": {
"type": "string",
"description": "The identifier of the order to retrieve.",
}
},
"required": ["order_id"],
},
},
}
]
response = client.chat.completions.create(
model="<ORG>/<MODEL_NAME>",
messages=[
{
"role": "user",
"content": "Check the current processing status of order A-1024.",
}
],
tools=tools,
tool_choice="auto",
temperature=1.0,
top_p=0.95,
max_tokens=32768,
extra_body={
"top_k": 20,
},
)
message = response.choices[0].message
if message.tool_calls:
for tool_call in message.tool_calls:
print("Tool name:", tool_call.function.name)
print("Tool arguments:", tool_call.function.arguments)
else:
print(message.content)
Direct Inference with Transformers
For local testing, Occamy-35B can also be loaded directly with Hugging Face Transformers.
Install the required dependencies:
pip install -U transformers accelerate pillow torchvision
Run inference:
from transformers import AutoModelForMultimodalLM, AutoProcessor
model_id = "<ORG>/<MODEL_NAME>"
processor = AutoProcessor.from_pretrained(
model_id,
trust_remote_code=True,
)
model = AutoModelForMultimodalLM.from_pretrained(
model_id,
torch_dtype="auto",
device_map="auto",
trust_remote_code=True,
)
messages = [
{
"role": "user",
"content": [
{
"type": "text",
"text": "Analyze this problem and provide a complete solution.",
}
],
}
]
inputs = processor.apply_chat_template(
messages,
add_generation_prompt=True,
tokenize=True,
return_dict=True,
return_tensors="pt",
).to(model.device)
outputs = model.generate(
**inputs,
max_new_tokens=8192,
)
generated_ids = outputs[0][inputs["input_ids"].shape[-1] :]
response = processor.decode(
generated_ids,
skip_special_tokens=True,
)
print(response)
For production deployment, long-context inference, or high-concurrency workloads, we recommend using SGLang or vLLM instead of direct Transformers inference.
- Downloads last month
- 9