Ornith 1.0 35B¶
Current status and review date¶
Decision snapshot
- Product role: Historical agentic-coding specialist lead.
- Selected or best-qualified configuration: NGC vLLM 0.19 with compressed-tensors FP8 weights, FP8 KV cache, 131,072 served tokens, Qwen reasoning and tool parsers, and thinking disabled.
- Measured hardware: One RTX PRO 6000 on Primary Node.
- Evidence: Functional, capacity, and bounded quality evidence includes a 131K needle pass, 20/20 tool calls, a session pass, intelligence 1/2, and 29.2 tok/s.
- Decision:
no-promotion; retain as historical specialist evidence. - Important limitation: The checkpoint commit, executed image digest,
and complete source recipe were not retained, so the run is
historical-invalidfor exact reproduction. - Review dates: Retained evidence cutoff: 2026-07-10. Dossier-format review: 2026-08-31.
Review narrative¶
2026-07-10 — Specialist lead retained without promotion¶
Ornith passed the long-context needle and tool/session checks on one RTX PRO 6000, but its bounded intelligence slice passed only one of two prompts. The result remains useful as an agentic-coding specialist lead rather than a promotion result. Default thinking also exhausted small output budgets, so the tested lane disabled thinking. Vendor quality claims were not locally verified and played no role in the decision.
Immutable identity¶
Model¶
- Repository:
deepreinforce-ai/Ornith-1.0-35B-FP8 - Checkpoint commit: Not retained.
Runtime¶
- Engine release: NGC vLLM 0.19.
- Executed image digest and engine commit: Not retained.
The missing checkpoint and runtime pins make the exact-rerun identity
historical-invalid.
Tested hardware and topology¶
Measured lane¶
- Host label: Primary Node.
- Hardware: one RTX PRO 6000.
- Topology: single-card candidate serve.
No second hardware lane was tested.
Engine, quantization, KV, context, and concurrency recipe¶
Retained run settings¶
- Engine: NGC vLLM 0.19.
- Weight format: compressed-tensors FP8.
- KV cache: FP8.
- Served context: 131,072 tokens.
- Parsers: Qwen reasoning and tool parsers.
- Thinking: disabled.
- Concurrency: Not recorded in the dossier evidence.
Reconstruction boundary¶
The tracked experiment Compose file is the closest public reconstruction aid. The dated artifact did not retain a source recipe or launch command, so the Compose service must not be presented as proof of the exact command that produced the measurement.
Evidence by measurement class¶
Functional and capacity¶
- 131K needle: pass.
- Tool calls: 20/20.
- Session gate: pass.
Bounded quality and performance¶
- Intelligence slice: 1/2.
- Decode performance: 29.2 tok/s.
Evidence boundary¶
The result carries historical-invalid identity because the immutable model
and runtime pins are incomplete. Raw request and response artifacts were
not retained in the public dossier bundle.
Decision and promotion state¶
Retained¶
- Historical agentic-coding specialist lead.
Not authorized¶
no-promotion.- No route or deployment change is implied by this historical record.
Failures and gotchas¶
Serving behavior¶
- Default thinking exhausted small output budgets; the measured lane disabled thinking.
Evidence limitations¶
- Checkpoint commit, image digest, engine commit, source recipe, and exact launch command: Not retained.
- Vendor quality claims: Not tested locally.