Skip to content
CrucibleMark

Ornith 1.0 35B

What FP8 block quantization delivers with Ornith-1.0-35B-FP8: an Open Weights MoE with only around 3 of 35 billion active parameters per token runs on a single GPU and brings 262,144 tokens of context, native thinking, and tool calling. DeepReinforce trained the model to learn its own agentic approach rather than working with a fixed rule set. MIT license, commercial use, and fine-tuning without restrictions.

DeepReinforce Version 1.0 Commercial use permitted MoE 35 B (3 B active) 262 K Context 05/2026 locally tested

  • Open Weights
  • Workstation
  • vLLM
  • Text
  • Batch

Sovereign Risk: LOW DeepReinforce is a US-based RL research organization. The model is available on Hugging Face under the MIT license without regional restrictions (79,608 downloads/month). Lineage: Qwen3.5-35B-A3B (hybrid MoE base, Alibaba Cloud) + Gemma 4 → DeepReinforce Ornith-1.0-35B (RL post-training) → official FP8 block quantization (E4M3) by the same author. No Chinese NSL risk, no US CLOUD Act risk when operated locally, as it is a pure Open Weights model with no cloud API requirement.

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

Name P1 P2 Combined
EU License Research 100 40 64
HTTP Fetch & Extract 80 80 80
Tool Failure Handling (404) 40 40 37
Web Search & Tool Selection 35 20 28.3
URL Construction & Fetch 80 80 80
Multilingual Search & Synthesis 100 80 90

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

Reliability

  • Tool Call Valid: No
  • Retry: Not required
  • Hallucination: Not 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
6.12
First Request
MCP
1.02
Protocol Latency
Synthesis
48.09
Response Generation
Total
331.4
Sum of All Phases
Token
22902
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

Updated

Deployment Verdict

Conditional deploy, because tool execution is often usable, but the model does not work reliably enough in tool selection and result synthesis for autonomous MCP pipelines. The absence of hallucination findings mitigates the risk; the invalid tool calls and the only moderate overall impression do not.

Tool Execution Profile

Ornith 1.0 35B can execute tools, but not with consistent protocol fidelity. P1 of 80.83 indicates basic operational capability. The problem lies not in simple retrieval but in selecting the right tool. On the Web Search & Tool Selection test — which checks whether the model chooses search over fetch without being prompted — the model clearly falls short with P1 35. This argues against robust tool intelligence and suggests a fixed pattern instead: known URLs or direct fetch paths work; open research paths perform noticeably worse. The URL Construction & Fetch test, which evaluates deriving a target URL from internal knowledge, confirms this with P1 80. HTTP Fetch & Extract and Multilingual Search & Synthesis also show that the model processes available sources adequately when the access path is already clear. For dynamic tool routers, that is too weak. For pre-structured pipelines with a narrow tool selection, it is usable.

Synthesis Fidelity

How well does it consolidate tool results? Only to a limited degree. P2 of 62.50 is the actual bottleneck. Ornith extracts relevant facts often enough, but does not consistently condense them into precise, decision-ready answers. This is most visible in EU License Research with P2 40 — even though the tool access itself succeeds — and in Tool Failure Handling (404), also with P2 40.

Does it stay within the tool result or fall back on training data? Here the trust verdict is better than the synthesis quality. In the honeypot EU License Research — which checks whether current license restrictions are actually answered from web sources rather than training knowledge — no hallucination was detected. That is a strong signal for compliance-adjacent pipelines. It does not mean the answer quality is high. It only means the model does not undermine the infrastructure with fabricated source content.

Error Resilience

On the 404 test, which measures transparent handling of failed tool calls, Ornith does not hallucinate substitute content. That is a productively relevant positive. However, the communication of the error remains too weakly synthesized and is not always decision-oriented. Acceptable for supervised systems. Too unreliable for fully autonomous agent loops.

Sovereignty Profile

Locally deployable, MIT-licensed, and usable without cloud dependency. At the same time, it sits -1.85 points below the fleet average of 67.58. That is close enough to the fleet average to be seriously relevant for sovereign environments, but not strong enough to offset quality deficits through operational freedom.

Conclusion & Recommendation

Suitable for local, sovereign MCP pipelines with fixed tool paths, human sign-off, and clear fallbacks. Not suitable as an autonomous orchestrator that must reliably choose between search, fetch, and error handling on its own. If you are looking for a local model for document-adjacent research, multilingual synthesis, and controlled tool use, Ornith can deliver. If the pipeline demands independent tool selection and robust final synthesis, you should deploy it only behind guardrails and with a supervisor in place.

This analysis was generated automatically based on benchmark data. Model used: GPT 4.5 by OpenAI. The raw data and full methodology are documented in the GitHub project.