Add Occamy-1.0 model card

Document model scope, training recipe, evaluation results, cost efficiency, serving, and limitations.

README.md

@@ -1,3 +1,210 @@
 ---
 license: apache-2.0
+library_name: transformers
+pipeline_tag: image-text-to-text
+base_model: Qwen/Qwen3.6-35B-A3B
+tags:
+- agent
+- agentic
+- co-work
+- tool-use
+- long-context
+- mixture-of-experts
+- coding
 ---
+
+<div align="center">
+
+# Occamy-1.0
+
+**Open Pareto-frontier 35B Intelligence for Co-work**
+
+[Model weights](https://huggingface.co/Accio-Lab/Occamy-1.0) · [GitHub](https://github.com/Accio-Lab/occamy) · [Training infrastructure](https://github.com/Accio-Lab/Dressage)
+
+</div>
+
+## Overview
+
+Occamy-1.0 is an open, cost-efficient co-work model developed by the Accio Team. It is built by further post-training [Qwen3.6-35B-A3B](https://huggingface.co/Qwen/Qwen3.6-35B-A3B) for long-horizon, stateful agent workloads across terminals, files, structured APIs, and productivity tools.
+
+Co-work tasks can invoke a model dozens or hundreds of times. Their practical cost therefore depends on the complete episode, not only on the price of a single response. Occamy-1.0 is designed to deliver reliable task execution at a compact 35B-A3B scale. Across our co-work evaluation suite, it is consistently among the strongest comparably sized models and remains competitive with substantially larger systems on several tasks. Under the evaluation and pricing protocol described in our technical report, Occamy lies near the low-cost knee of the observed cost-performance Pareto frontier.
+
+## Highlights
+
+- **Built for co-work:** optimized for multi-step, stateful work rather than isolated question answering.
+- **Long-horizon execution:** trained to preserve progress across tool calls, failures, context compaction, pruning, and other harness-level history rewrites.
+- **Broad agentic capability:** strong co-work performance while retaining competitive tool-calling, coding, and instruction-following ability.
+- **Cost-efficient serving:** 35B total parameters with 3B activated per token, a practical operating point for self-hosted agent systems.
+- **Open ecosystem:** Apache-2.0 model weights, an open GitHub repository, and an open-source version of our multi-harness training infrastructure.
+
+## Model details
+
+| Item | Value |
+|---|---|
+| Developer | Accio Team |
+| Base checkpoint | Qwen3.6-35B-A3B |
+| Architecture | Mixture-of-Experts, 35B total / 3B activated |
+| Context length | 262,144 tokens |
+| Precision | BF16 |
+| Primary focus | Long-horizon co-work and tool-using agents |
+| License | Apache-2.0 |
+
+The checkpoint retains the Qwen3.6 multimodal architecture, but the current Occamy release and reported evaluation focus on text- and tool-based agent execution. Browser and GUI interaction are not currently supported or claimed.
+
+## Training overview
+
+Occamy-1.0 uses a specialization-and-consolidation recipe rather than training from a base model:
+
+1. **Marathon Expert:** full-parameter SFT followed by Hierarchical Decoupled Policy Optimization (HDPO), targeting sustained execution over long, stateful episodes.
+2. **Sprint Expert:** full-parameter SFT over a broader distribution of shorter-horizon agentic workloads, including coding, information gathering, and structured tool use.
+3. **Capability harmonization:** model merging combines the complementary experts, followed by Single-Rollout Asynchronous Optimization (SAO) on a broad co-work task mixture.
+
+Training data are execution-grounded: tasks are paired with runnable environments, observable state transitions, realized tool behavior, trajectory evidence, and task-level grading. Multi-harness adapters preserve harness-specific token, tool, rewrite, and state semantics while exposing a common trajectory and replay contract to the learner.
+
+[Dressage](https://github.com/Accio-Lab/Dressage) is the open-source release of the infrastructure family used for multi-harness agentic training, token-exact trajectory capture, and state replay.
+
+## Evaluation
+
+The table below summarizes the current Occamy-1.0 results. Exact harness versions, budgets, sampling settings, judge models, and comparison rules are documented in the technical report.
+
+| Category | Benchmark | Metric | Occamy-1.0 |
+|---|---|---:|---:|
+| Co-work | Claw-Eval | Average | 82.20 |
+| Co-work | Claw-Eval | Pass³ | 71.40 |
+| Co-work | WildClawBench | avg@3 | 49.16 |
+| Co-work | CommerceAgentBench | Pass¹ | 37.38 |
+| Co-work | Business Arena | Avg. final net worth | $79,868 |
+| Co-work | GDPval | Score | 1128 |
+| Co-work | OfficeQA Pro | Accuracy | 48.10 |
+| Co-work | τ³-Bench (Banking) | Pass¹ | 37.10 |
+| Tool calling | AutomationBench | Pass¹ / Partial | 27.60 / 69.10 |
+| Tool calling | BFCL v4 | Score | 65.40 |
+| Tool calling | VitaBench | Score | 41.75 |
+| Coding | Terminal-Bench 2.1 | Accuracy | 59.00 |
+| Instruction following | IFEval | Score | 91.53 |
+
+We do not currently report a standalone search benchmark. Search and information gathering may appear inside co-work tasks, but they should not be interpreted as a separate, controlled search-capability evaluation.
+
+### Efficiency and reliability
+
+On the combined Claw-Eval T/C tasks, Occamy improves task success while using less interaction and producing fewer execution failures than its starting checkpoint under the same protocol.
+
+| Metric | Qwen3.6-35B-A3B | Occamy-1.0 |
+|---|---:|---:|
+| Claw-Eval average | 69.5 | **82.2** |
+| Trial success rate | 62.81% | **77.55%** |
+| Tokens per trajectory | 185,999 | **149,713** |
+| Tool calls per trajectory | 14.24 | **12.07** |
+| Trace wall time | 74.58 s | **39.96 s** |
+| Timeout rate | 9.88% | **2.18%** |
+
+For the aggregate cost-performance analysis, Claw-Eval, WildClawBench, AutomationBench, and GDPval are min-max normalized across the compared systems and weighted equally. Per-task cost is computed from measured token usage using a common input, cache-read, and output pricing protocol. This is an inference-cost estimate for comparison, not a complete deployment total-cost-of-ownership analysis.
+
+## Quickstart
+
+Occamy-1.0 uses the supplied <code>chat_template.jinja</code>. We recommend serving it behind an OpenAI-compatible endpoint with a recent version of SGLang or vLLM.
+
+### SGLang
+
+The following example uses tensor parallelism across eight GPUs and enables Qwen reasoning and tool-call parsing:
+
+~~~bash
+uv pip install "sglang[all]>=0.5.10"
+
+python -m sglang.launch_server \
+  --model-path Accio-Lab/Occamy-1.0 \
+  --port 8000 \
+  --tp-size 8 \
+  --mem-fraction-static 0.8 \
+  --context-length 262144 \
+  --reasoning-parser qwen3 \
+  --tool-call-parser qwen3_coder
+~~~
+
+### vLLM
+
+~~~bash
+uv pip install "vllm>=0.19.0" --torch-backend=auto
+
+vllm serve Accio-Lab/Occamy-1.0 \
+  --port 8000 \
+  --tensor-parallel-size 8 \
+  --max-model-len 262144 \
+  --reasoning-parser qwen3 \
+  --enable-auto-tool-choice \
+  --tool-call-parser qwen3_coder
+~~~
+
+If memory is limited, reduce the context length. Long-horizon agent performance can depend on the available context and the harness's compaction policy, so results may differ from the reported configuration.
+
+### OpenAI-compatible client
+
+~~~python
+from openai import OpenAI
+
+client = OpenAI(
+    base_url="http://localhost:8000/v1",
+    api_key="EMPTY",
+)
+
+response = client.chat.completions.create(
+    model="Accio-Lab/Occamy-1.0",
+    messages=[
+        {
+            "role": "user",
+            "content": "Inspect the workspace, fix the failing tests, and summarize the changes.",
+        }
+    ],
+    max_tokens=8192,
+    temperature=1.0,
+    top_p=0.95,
+    extra_body={"top_k": 20},
+)
+
+print(response.choices[0].message.content)
+~~~
+
+For agent deployment, provide tool schemas through the serving API and let the harness own environment state, timeouts, retries, and history compaction. Replacing the supplied template or parser can change tool-call behavior.
+
+## Intended use
+
+Occamy-1.0 is intended for research and development of:
+
+- terminal- and file-based co-work agents;
+- structured API and productivity-tool workflows;
+- long-horizon task execution with stateful environments;
+- agentic post-training, replay, and evaluation;
+- coding and repository-level assistance inside sandboxed harnesses.
+
+Use sandboxing, least-privilege credentials, action validation, and human confirmation for consequential operations.
+
+## Limitations
+
+- Performance is sensitive to the harness, tool schemas, timeout budget, retry policy, context length, and history-rewrite behavior.
+- Browser and GUI interaction are not currently supported or evaluated.
+- The model may issue invalid calls, misread environment state, repeat actions, or stop before completing a task.
+- Benchmark scores obtained with different harnesses or budgets are not directly comparable.
+- The reported cost analysis estimates inference cost under a common pricing protocol; it does not include complete infrastructure, engineering, or operational costs.
+- This release does not provide a safety guarantee for autonomous use in high-impact domains. Users are responsible for task-specific evaluation and safeguards.
+
+## Resources
+
+- **Model:** https://huggingface.co/Accio-Lab/Occamy-1.0
+- **GitHub:** https://github.com/Accio-Lab/occamy
+- **Training infrastructure:** https://github.com/Accio-Lab/Dressage
+- **Technical report and selected training data:** links will be added with the public release.
+
+## Citation
+
+~~~bibtex
+@misc{accio2026occamy,
+  title        = {Occamy-1.0: Open Pareto-frontier 35B Intelligence for Co-work},
+  author       = {{Accio Team}},
+  year         = {2026},
+  howpublished = {\url{https://huggingface.co/Accio-Lab/Occamy-1.0}}
+}
+~~~
+
+## License
+
+Occamy-1.0 is released under the Apache License 2.0. Please also review the license and acceptable-use terms of the base model.