Bounded greedy parity checks; final MTP3 speed unmeasured
Sizes are decimal GB of weight files, not peak RAM/VRAM. GGUF image input requires the separate 0.90 GB F16 vision projector. KV cache, context length, concurrency and runtime buffers need additional memory. The MLX collection groups the Apple Silicon candidates.
BF16, FP8 and NVFP4 each passed a nine-case vLLM compatibility check on a single H200, covering text, code, JSON, tool calls and follow-up, and images. NVFP4 used Marlin W4A16. These checks used ordinary decoding; the separate MTP head remains experimental. Format-specific sample checks do not establish equal benchmark quality or a general speed ranking.
Occamy-1.0 is a compact agentic model purpose-built for real-world co-work: long-horizon, stateful tasks that require coordinated use of search, code, tools, files, structured APIs, and productivity software. Starting from the post-trained Qwen3.6-35B-A3B checkpoint, Occamy concentrates further training on reliable execution, persistent state tracking, recovery, and follow-through rather than relearning general capabilities from scratch.
Co-work specialization: Designed for sustained execution across multi-step professional workflows, not isolated question answering.
Compact inference footprint: A 35B-total, 3B-active Mixture-of-Experts model that keeps long-running agent workloads practical.
Long-horizon continuity: Designed to keep work coherent across tool calls, delegated runs, and history rewrites such as context compaction.
Broad agentic capability: Co-work gains are accompanied by strong tool calling, terminal coding, and instruction following.
Execution-grounded training: Supervised fine-tuning spans general agentic work, long-horizon interaction, software engineering, and tool-call grounding.
Open training stack: The multi-harness reinforcement-learning infrastructure used to train Occamy is released as Dressage.
Note. Occamy is optimized for common co-work workloads, not as a replacement for frontier models on every task. Retrieval-heavy and simulated-user tasks still have headroom, and native browser or desktop visual interaction is not part of the current co-work training interface.
| Post-training | Full-parameter SFT, HDPO, model merging, and SAO
Architecture fields follow the starting checkpoint's published model card. Occamy post-trains the language backbone without changing the architecture; the vision encoder and projector are frozen during SFT. The released checkpoint configuration remains the source of truth for serving limits.
Within each size group, bold denotes the best result and underlining denotes the second-best result. * Official model-card or Artificial Analysis result. † Reproduced on the public task release.
Across Claw-Eval, WildClawBench, AutomationBench, and GDPval, Occamy-1.0 lies near the low-cost knee of the empirical Pareto frontier. Relative to its Qwen3.6-35B-A3B starting checkpoint, it delivers a large aggregate capability gain with only a modest change in measured per-task inference cost. Benchmark scores are equally weighted after per-benchmark min-max normalization, and costs are macro-averaged per task under the frozen pricing protocol used in the report.
The Marathon Expert learns sustained execution and accuracy-conditioned efficiency, while the Sprint Expert preserves broader agentic capability. A uniform parameter-space merge combines both experts into one checkpoint with no inference-time routing or ensembling, and a final Single-Rollout Asynchronous Optimization (SAO) stage refines the merged policy on a broad co-work mixture.
The deduplicated SFT union across both experts is:
Data source
Trajectories
Average length
Tokens
General agentic
5,418
37.7K
204.1M
Long-horizon interactive agents
923
95.8K
88.4M
Terminal and software engineering
1,228
35.1K
43.1M
Tool-call grounding
7,429
9.1K
67.7M
Overall
14,998
26.9K
403.3M
Training tasks are grounded in executable environments with observable state transitions and task-level grading. The open-source Dressage stack provides multi-harness execution, token-exact trajectory capture, sandbox integration, and multi-segment conversion for reinforcement learning.
Occamy-1.0 keeps the Qwen3.6-35B-A3B architecture, so the upstream deployment recipe is the reference serving path. The examples below mirror that recipe with eight-way tensor parallelism and its full context length; adjust both to fit your hardware and confirm them against the released Occamy checkpoint configuration.
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 this repository, fix the failing test, and explain the change.",
}
],
max_tokens=32768,
temperature=1.0,
top_p=0.95,
presence_penalty=1.5,
extra_body={
"top_k": 20,
"chat_template_kwargs": {
"enable_thinking": True,
"preserve_thinking": True,
},
},
)
print(response.choices[0].message.content)
For multi-turn agent runs, retain the complete assistant message returned by the server, including reasoning content and tool calls, then append tool results using the standard OpenAI chat-completions schema. This preserves the execution context that Occamy relies on across long workflows.
Occamy was trained and evaluated across multiple harnesses, including OpenClaw, Hermes Agent, and Accio Work. It can be integrated with other tool-using agent frameworks through the same OpenAI-compatible API.
