Occamy-1.0 is a co-work model from Alibaba’s Accio Team, further trained from the post-trained Qwen3.6-35B-A3B checkpoint. This 35A3B model carries out workflows spanning search, tools, code, and files, with an emphasis on state tracking, recovery, and verified completion.
My contribution
Core contributor to Occamy-1.0 across agentic post-training algorithms, training infrastructure, and data. My work includes:
- Co-designed the staged SFT/HDPO, model-soup, and SAO post-training recipe, with trainable context compaction and cross-segment credit assignment.
- Architected multi-harness training infrastructure connecting sandboxed OpenClaw, Hermes, and Accio Work execution to asynchronous actor–critic learning through a shared trajectory, state, and reward contract.
- Designed execution-grounded data construction: environment-first and capability-first synthesis, executable task contracts, state-aware graders, verified training admission, and coverage-controlled mixtures. The team’s deduplicated SFT union spans 14,998 trajectories and 403.3M tokens.
- Preserved token-exact proxy/TiTO replay, environment-state reconstruction, segment lineage, and loss masks as supporting guarantees for the training path.
- Matched evaluations of task success, token use, latency, and tool failures.
Training recipe
The team develops two complementary checkpoints from the same starting model. The Marathon Expert uses full-parameter SFT followed by HDPO for sustained, long-horizon execution. The Sprint Expert uses SFT on a broader distribution of shorter tool-use, coding, and search tasks. A uniform model soup merges the two, followed by SAO reinforcement learning on the merged policy’s interaction distribution. GRPO is a comparison objective, not the final post-merge training stage.
The deduplicated SFT union contains 14,998 trajectories and 403.3M tokens across general agentic work, long-horizon interaction, software engineering, and tool-call grounding.
Results
The latest report’s model-level results, compared with Qwen3.6-35B-A3B:
| Benchmark | Qwen3.6-35B-A3B | Occamy-1.0 |
|---|---|---|
| Claw-Eval average score | 69.5 | 82.2 |
| Claw-Eval Pass³ | 54.8 | 71.4 |
| WildClawBench | 40.40 | 49.16 |
| AutomationBench strict pass rate | 7.5% | 27.6% |
| AutomationBench partial score | 39.4 | 69.1 |
| CommerceAgentBench | 19.6 | 37.40 |
| Business Arena | $44,751 | $79,868 |
| GDPval | 1,004 | 1,128 |
| OfficeQA Pro | 39.1 | 48.1 |
| τ³-Bench (Banking) | 11.9 | 37.1 |
| BFCL v4 | 63.19 | 65.40 |
| VitaBench | 34.25 | 41.75 |
| Terminal-Bench 2.1 | 49.5 | 59.0 |
| IFEval | 86.90 | 91.53 |
Each row uses the benchmark’s own score or unit; these are not interchangeable accuracy measures. Native browser or desktop visual interaction is not part of the current co-work training interface.
Cost–performance frontier
Occamy reaches the low-cost knee of the observed cost–performance Pareto frontier in the report’s comparison across four co-work benchmarks.
Cost is first averaged per task within each benchmark, then equally averaged across benchmarks; 35B-A3B models use a common pricing schedule for comparison. This is a standardized cost comparison, not a measured GPU bill.
Updated September 9, 2026 from the official technical report and project website. Results describe the team’s model; the contribution summary above describes my work.