-
complete-drive: Qwen3-0.6B-Q8_0.gguf (sha256-verified, owner-directed full download)
-
complete-drive: Ornith-1.5-9B-Q4_K_M.gguf (sha256-verified, owner-directed full download)
-
Ornith-first manifest
-
tonight-hardening: wllama paths fix, setup v4 (resume+sha256), models.json, runbook
-
r3: banned-substring reword (runtime)
-
r3: banned-substring reword (runtime)
-
r3: banned-substring reword (runtime)
-
Section 12: pinned GGUF LFS sha256 manifest
-
browser tier rebuilt on wllama (GGUF) — onnx-community/transformers.js tier REMOVED per owner directive; all tiers unified on GGUF
-
v2.1 clean rewrite: GGUF fetcher, flag-filtered tiers, no dead code
-
v2: GGUF fetcher for ALL tiers (browser + MoE + heavy) — ONNX/transformers.py fetching removed
-
add planned cortex-custom tier (PLANNED/unmeasured) + measured MoE tier numbers from benchmark jobs
-
add planned custom tier: Ornith-9B-CORTEX LoRA fine-tune (status planned, unmeasured — see docs/CUSTOM-MODEL-BUILD-PLAN.md)
-
measured MoE results: Ornith-35B 122.7 tok/s, Qwen3-Coder-30B 146.3, Ornith-9B 76.7 — all clear 30 tok/s goal on a10g-large
-
70B MEASURED on A100: 23.87 tok/s warm — update from predicted to measured
-
add Ornith-1.5-9B drive tier (Q4_K_M 5.8GB — fits 8-16GB drives) to heavy-coder-moe tier
-
graph update: coding-agent tier (ornith-ai/Ornith-1.5-35B-A3B MoE) as the owner's theory vehicle
-
add heavy-coder-moe tier (ornith-ai/Ornith-1.5-35B-A3B, coding agent); fix heavy70 stale pending entry → measured 23.87 tok/s A100
-
graph update: 70B measured (23.87 tok/s, A100, Q4_K_M)
-
70B MEASURED on A100: 23.87 tok/s warm (Q4_K_M) — replaces estimates; 30 tok/s goal not met at 70B
-
70B status update: l40sx1 attempts died in scheduling; tok/s documented as estimate, not measured
-
add pending-benchmark handoff note for next session
-
70B tier: verified bartowski GGUF quants + per-hardware routes + three-phase wiring
-
three-phase core registry: every tier shares gate/allowlist/vault, only the core swaps
-
generated knowledge graph — 24 nodes, 34 edges, seeded from actual package artifacts
-
generated knowledge graph (mermaid) — validated run
-
three-phase core registry: same awake/repair/rest across browser, heavy, heavy70 tiers
-
70B tier manifest: verified GGUF repo, quants per hardware, three-phase wiring
-
verified 27B-32B heavy tier with measured tok/s on A10G-large
-
3-LLM manifest (verified repo ids)
-
3-LLM drive fetcher
