Occamy 1.0 35B-A3B (Accio-Lab) (Thinking)

Occamy 1.0 by Accio-Lab is an agentic derivative of the Qwen3.6-35B-A3B checkpoint, focused on long-horizon co-work sessions with tools, structured APIs, and persistent state tracking. The NVFP4 quantization is selective: only the routed experts are quantized, while attention, router, embeddings, and output head remain in BF16. The 35-billion-parameter MoE activates only 3 billion parameters per token and supports 262,000 tokens of context. Apache 2.0 license and documented provenance with recipe, data, and validation artifacts.

Accio-Lab Version 1.0 Commercial use permitted MoE 35 B (3 B active) 262 K Context 12/2025 locally tested

  • Open Weights
  • Workstation
  • vLLM
  • Text
  • Vision
  • Unusable

Sovereign Risk: MEDIUM Occamy 1.0 NVFP4 is a community derivative of Qwen/Qwen3.6-35B-A3B, with a published provenance trail including the quantization recipe, data-provenance file, and validation artifacts. The upstream base is Apache 2.0, the checkpoint runs locally, and the NVFP4 export is limited to routed experts, but Accio-Lab’s organizational jurisdiction is not publicly documented, so the provenance risk remains medium.

Tool-use profile: 6 assets in detail

Comparison of asset performance (P1/P2/Combined) against the fleet average

Asset performance (radar)

Score Breakdown vs. Fleet Average


Tool-use details

Asset performance, reliability, and runtime profile

CrucibleMark evaluates tool use across 6 independent tests. Click a test name for details.

Reliability

  • Tool Call Valid: No
  • Retry: Not required
  • Hallucination: Detected

Reliability measures how consistently a model actually executes tool calls: Tool Call Valid schema and format accepted, Retry Required successful only after retry, Hallucination Flag fabricated tools or parameters detected. All three green means production-ready.

Runtime profile

Call 1
4.02
First Request
MCP
1.32
Protocol Latency
Synthesis
136.43
Response Generation
Total
850.61
Sum of All Phases
Token
10986
Input + Output
Cost
$0
Cost per Run

The runtime profile shows the latency and cost metrics for the model run: Call 1 First Request, MCP Protocol Latency, Synthesis Response Generation, Total sum of all phases, supplemented by Token input and output and Cost cost per run.

Tool-use review

Created

Deployment Verdict

Conditional deploy, because tool execution is strong, but tool calls are not consistently valid and synthesis reliability falls short of what unattended production pipelines require, given the hallucination finding.

Tool Execution Profile

Occamy demonstrates clear tool intelligence rather than mere pattern-matching. In the Web Search & Tool Selection test — which checks whether the model selects web_search over fetch without being prompted — it makes the correct decision consistently. This points to usable planning logic in MCP-backed workflows. EU License Research also runs cleanly on the P1 side, which matters because the model must actively retrieve current web sources there.

Performance weakens on precision work after the decision point. In the URL Construction test — which checks for the correct target URL from internal knowledge and the subsequent fetch — execution is merely adequate, not deterministic. That is exactly where the production risk lies: the model often knows which tool is needed, but the concrete call is not stable enough. The “Tool call valid: false” finding therefore weighs heavier than the solid overall P1 impression. At least no retry was required. This reads more like an execution and precision problem than a protocol comprehension failure.

Synthesis Fidelity

How well does it consolidate tool results? Only to a limited degree. The P2 score of 42.50 is the clear weak point. Multilingual Search & Synthesis is particularly notable: the model finds information across language boundaries but consolidates it imprecisely and loses relevance in the process. Quality also breaks down noticeably in URL Construction & Fetch when merging retrieved content. For pipelines requiring exact extraction, compliance citations, or reliable result consolidation, this is insufficient.

Does it stay within tool output or fall back on training data? In the Honeypot EU License Research test it formally stays in safe territory: no hallucination, meaning no fallback to stale training knowledge despite a current licensing question. That is the most important trust anchor. The global hallucination finding nonetheless remains a security risk. Once a model can output fabricated facts as tool results, it undermines confidence in the entire tool infrastructure.

Error Resilience

On Tool Failure Handling (404) — which tests for transparent handling of a failed retrieval — Occamy behaves in a production-appropriate manner. It does not hallucinate page content and communicates the failure in a comprehensible way. This behavior is acceptable in robust pipelines precisely because the orchestrator can then escalate or replan cleanly.

Sovereignty Profile

Locally deployable, open-weight, and therefore attractive for sovereign deployments. Combined 70.00, which is 1.83 points above the fleet average of 68.17. Competence is thus fleet-viable, but not without guardrails.

Summary & Recommendation

Suitable for local, sovereign agent pipelines with strong external validation, well-defined tool schemas, and downstream result verification. Well suited for research orchestration, search-and-fetch chains, and fault-tolerant assistance workflows. Not suitable for high-trust pipelines without human or programmatic oversight — particularly for compliance, multilingual consolidation, and any workflow in which the synthesis itself must qualify as a reliable end product.

This analysis was generated automatically based on benchmark data. Model used: GPT-5.4 by OpenAI. The raw data and full methodology are documented in the GitHub project.