coulomb/freedom-intelligence
4
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Register reserve under agents hub and pin VAULT storage.

Browse source 
Classify freedom-intelligence for State Hub (agents domain), sync FI-WP-0001..0003 with hub IDs, enrich catalog entries with profile/SWOT, and pin the open-weight reserve to D:\vault\coulomb\freedom-intelligence\ with strategic S-tier policy.
This commit is contained in:
tegwick 2026-07-28 00:25:21 +02:00
parent 00f469662b
commit a83ef0a79a
29 changed files with 1343 additions and 279 deletions
Download patch file Download diff file Expand all files Collapse all files

47
inventory/catalog/BAAI__bge-m3__candidate.yaml
View file

@ -7,6 +7,7 @@ source:
url: https://huggingface.co/BAAI/bge-m3
revision: main
model_card_url: https://huggingface.co/BAAI/bge-m3
project_url: https://github.com/FlagOpen/FlagEmbedding
license:
spdx: MIT
url: https://huggingface.co/BAAI/bge-m3
@ -14,6 +15,41 @@ license:
allows_local_ops: true
allows_fine_tune: true
notes: ""
profile:
summary: "Default multilingual dense embedding model for local RAG."
original_source: https://huggingface.co/BAAI/bge-m3
use_cases:
- "NetKingdom / Coulomb document and wiki retrieval"
- "Multilingual (DE/EN/…) semantic search"
- "Hybrid retrieval experiments (dense + multi-granularity features)"
- "Offline RAG in sandboxed agents"
- "Indexing ops runbooks and code comments"
sweet_spots:
- "Self-hosted production-quality multilingual embed"
- "MIT license; easy ops story"
- "Strong general retrieval without API spend"
- "Pairs cleanly with a small cross-encoder reranker"
not_ideal_for:
- "Generation / chat (not an LLM)"
- "Pure code retrieval if a code-specialized embed clearly wins A/B"
- "Ultra-tiny edge when EmbeddingGemma-class is enough"
capability_notes: >
BAAI BGE-M3 is a staple open embedder: multi-lingual, multi-granularity,
widely deployed for RAG. Primary R-tier retrieval backbone for the lab.
swot:
strengths:
- "Proven multilingual retrieval quality for self-host"
- "MIT; small enough for CPU/GPU flexibility"
- "Mature FlagEmbedding ecosystem"
weaknesses:
- "Not optimized solely for code or for extreme long-context embed niches"
- "Newer embed lines (Qwen/Gemma/Nomic) may win specific A/B tests"
opportunities:
- "Domain-adapted embed fine-tune on Coulomb corpora later"
- "Companion bge-reranker for precision@k gains"
threats:
- "Leaderboard churn; risk of holding a stale 'default' without re-eval"
- "Pipeline lock-in if vector DBs assume fixed dimension without migration plan"
size:
total_bytes: 0
total_human: "~2 GB"
@ -31,12 +67,17 @@ collection:
storage_path: ""
brief_refs:
- research/2026-07-24-baseline-field-survey.md
reason: "P0 multilingual embedding staple for local RAG (docs, ops notes, DE/EN)."
tags: [embedding, multilingual, rag]
companions: []
reason: "P0/R multilingual embedding staple for local RAG (docs, ops notes, DE/EN)."
tags: [tier-r, embedding, multilingual, rag]
companions:
- BAAI__bge-reranker-v2-m3__candidate
notes: "Small download — within agent auto-collect band after license check once storage pinned."
history:
- at: "2026-07-24"
event: nominated
by: baseline-survey
detail: "P0 recommendation from initial deep research."
- at: "2026-07-24"
event: profile_swot_added
by: grok
detail: "schema 0.2 profile + SWOT."

42
inventory/catalog/BAAI__bge-reranker-v2-m3__candidate.yaml
View file

@ -7,6 +7,7 @@ source:
url: https://huggingface.co/BAAI/bge-reranker-v2-m3
revision: main
model_card_url: https://huggingface.co/BAAI/bge-reranker-v2-m3
project_url: https://github.com/FlagOpen/FlagEmbedding
license:
spdx: Apache-2.0
url: https://huggingface.co/BAAI/bge-reranker-v2-m3
@ -14,6 +15,39 @@ license:
allows_local_ops: true
allows_fine_tune: true
notes: "Confirm SPDX on card."
profile:
summary: "Cross-encoder reranker companion to BGE-M3 for higher-precision RAG."
original_source: https://huggingface.co/BAAI/bge-reranker-v2-m3
use_cases:
- "Second-stage rerank of top-k chunks from BGE-M3 (or hybrid) retrieval"
- "Improving answer grounding for NetKingdom / ops runbook Q&A"
- "Eval A/B of retrieval quality without changing the generator"
- "Low-cost precision boost on multilingual corpora"
sweet_spots:
- "Cheap RAG quality win after a solid bi-encoder"
- "Small VRAM; easy always-on sidecar"
- "Same BAAI/FlagEmbedding family as BGE-M3"
not_ideal_for:
- "First-stage retrieval over full corpora (too slow as sole retriever)"
- "Generation / chat"
- "When latency budget forbids a second pass"
capability_notes: >
Lightweight multilingual cross-encoder from the BGE family. W-tier optional
after R spine; high value per byte when RAG is a primary workload.
swot:
strengths:
- "Material precision lift for little disk/VRAM"
- "Proven pairing with BGE-M3"
- "Apache-friendly open stack"
weaknesses:
- "Cross-encoder cost scales with candidate count"
- "Not a substitute for better chunking or generators"
opportunities:
- "Default two-stage pipeline for lab RAG demos"
- "Tune candidate depth (k) per latency tier"
threats:
- "Newer BGE / open rerankers may supersede this checkpoint"
- "LLM-as-judge rerank experiments may reduce need for a dedicated model"
size:
total_bytes: 0
total_human: "~1–2 GB (estimate)"
@ -31,8 +65,8 @@ collection:
storage_path: ""
brief_refs:
- research/2026-07-24-baseline-field-survey.md
reason: "P1 — cheap RAG quality win as companion to BGE-M3."
tags: [reranker, rag]
reason: "P1/W — cheap RAG quality win as companion to BGE-M3."
tags: [tier-w, reranker, rag]
companions:
- BAAI__bge-m3__candidate
notes: ""
@ -41,3 +75,7 @@ history:
event: nominated
by: baseline-survey
detail: "P1 recommendation from initial deep research."
- at: "2026-07-28"
event: profile_swot_added
by: grok
detail: "schema 0.2 profile + SWOT."

46
inventory/catalog/Qwen__Qwen3-14B__candidate.yaml
View file

@ -7,6 +7,7 @@ source:
url: https://huggingface.co/Qwen/Qwen3-14B
revision: main
model_card_url: https://huggingface.co/Qwen/Qwen3-14B
project_url: https://qwenlm.github.io/
license:
spdx: Apache-2.0
url: https://huggingface.co/Qwen/Qwen3-14B
@ -14,6 +15,40 @@ license:
allows_local_ops: true
allows_fine_tune: true
notes: "Confirm card at download."
profile:
summary: "Mid-size Qwen3 step up from 8B for stronger single-GPU chat/code."
original_source: https://huggingface.co/Qwen/Qwen3-14B
use_cases:
- "Higher-quality local assistant when 8B is the bottleneck"
- "Harder coding and long-context drafting on T2 GPUs"
- "A/B baseline vs R1-Distill-14B (general vs reason-specialist)"
- "Domain FT when 8B capacity is insufficient"
sweet_spots:
- "Quality step within still-single-GPU open dense class"
- "Multilingual instruct continuity with Qwen3-8B"
- "Quota-friendly alternative to jumping to 70B"
not_ideal_for:
- "Always-on default if VRAM is tight (prefer 8B)"
- "Deepest reasoner tasks (prefer R1 distill or full R1 reserve)"
- "Embedding / retrieval"
capability_notes: >
W-tier optional after R spine. Collect when soft-quota headroom and a clear
quality gap vs Qwen3-8B show up in daily work. Prefer Instruct sibling if
separate at pin time.
swot:
strengths:
- "Clear capability bump over 8B without 70B cost"
- "Same Qwen3 stack and tooling as the R default"
- "Apache-friendly licensing typical"
weaknesses:
- "Near-duplicate niche vs strong 8B + selective 14B reason distill"
- "Still not frontier closed quality on hard agentic SWE"
opportunities:
- "Promote to R if lab defaults move off 8B"
- "FT target when domain data needs more capacity"
threats:
- "Disk spent better on S1 MoE or 70B dense under tight quota"
- "Next Qwen mid-size may obsolete this checkpoint quickly"
size:
total_bytes: 0
total_human: "~28 GB fp16 / ~9 GB Q4 (estimate)"
@ -31,12 +66,17 @@ collection:
storage_path: ""
brief_refs:
- research/2026-07-24-baseline-field-survey.md
reason: "P1 — stronger single-GPU chat/code when quota allows after P0."
tags: [instruct, qwen3]
companions: []
reason: "P1/W — stronger single-GPU chat/code when quota allows after R spine."
tags: [tier-w, instruct, qwen3]
companions:
- Qwen__Qwen3-8B__candidate
notes: ""
history:
- at: "2026-07-24"
event: nominated
by: baseline-survey
detail: "P1 recommendation from initial deep research."
- at: "2026-07-28"
event: profile_swot_added
by: grok
detail: "schema 0.2 profile + SWOT."

83
inventory/catalog/Qwen__Qwen3-72B__strategic.yaml Normal file
View file

@ -0,0 +1,83 @@
id: Qwen__Qwen3-72B__strategic
status: candidate
name: Qwen3-72B
org: Qwen
source:
kind: huggingface
url: https://huggingface.co/Qwen/Qwen3-72B
revision: main
model_card_url: https://huggingface.co/Qwen/Qwen3-72B
project_url: https://qwenlm.github.io/
license:
spdx: Apache-2.0
url: https://huggingface.co/Qwen/Qwen3-72B
allows_offline_retention: true
allows_local_ops: true
allows_fine_tune: true
notes: "Confirm card; use Instruct/Coder sibling if that is the capability peak for the line."
profile:
summary: "Strategic large Qwen dense — multilingual/tool open 70B-class reserve."
original_source: https://huggingface.co/Qwen/Qwen3-72B
use_cases:
- "Future multi-GPU dense instruct without MoE serving complexity"
- "Strong multilingual (DE/EN) offline assistant at 70B scale"
- "Domain FT base when 8B/14B capacity is too small"
- "Dense alternative to MoE giants for portable lab stacks"
sweet_spots:
- "S4 dense open with Qwen tooling continuity from the R 8B default"
- "Easier multi-GPU dense serve story than 600B MoE"
- "Apache-friendly licensing typical of Qwen3 line"
not_ideal_for:
- "Current default single-GPU ops"
- "Collecting both this and Llama-70B-class first under tight quota"
- "Embedding / retrieval"
capability_notes: >
Large dense strategic pick. Prefer Instruct/Coder sibling if that is the line
peak at pin time. If soft quota is tight after S1, pick at most one of
Qwen-72B vs Llama-70B-class first.
swot:
strengths:
- "High multilingual dense quality in open 70B class"
- "Same family as R-tier Qwen3-8B — transfer of prompts/tools"
- "Usually simpler ops than giant MoE"
weaknesses:
- "~40–150 GiB depending on quant; still heavy"
- "May lag top MoE on some frontier-open benches"
opportunities:
- "First dense strategic fill after R + optional S1 compressed"
- "FT / LoRA at 70B when facility upgrades"
threats:
- "Llama 70B / next Qwen large may be better use of same bytes"
- "Supersession by Qwen next large release"
size:
total_bytes: 0
total_human: "~145 GB fp16 / ~40 GB Q4 (estimate)"
hardware_class:
min_vram_gb_q4: 40
min_vram_gb_fp16: 145
notes: "T3+ to run; strategic dense multilingual open."
axes: [A, B, C]
priority: high
collection:
approved_by: ""
approved_at: null
downloaded_at: null
downloaded_by: ""
storage_path: ""
brief_refs:
- research/2026-07-24-nas-strategic-collection-plan.md
- docs/decisions/2026-07-24-nas-strategic-reserve.md
reason: "S4 strategic — large Qwen dense open for multilingual/tool capability reserve."
tags: [tier-s, strategic, dense, qwen3]
companions:
- Qwen__Qwen3-8B__candidate
notes: "Pick at most one of Llama-70B-class vs Qwen-72B-class first if quota tight after S1 MoE."
history:
- at: "2026-07-24"
event: nominated
by: operator-policy
detail: "Strategic dense open for NAS."
- at: "2026-07-28"
event: profile_swot_added
by: grok
detail: "schema 0.2 profile + SWOT."

45
inventory/catalog/Qwen__Qwen3-8B__candidate.yaml
View file

@ -7,6 +7,7 @@ source:
url: https://huggingface.co/Qwen/Qwen3-8B
revision: main
model_card_url: https://huggingface.co/Qwen/Qwen3-8B
project_url: https://qwenlm.github.io/
license:
spdx: Apache-2.0
url: https://huggingface.co/Qwen/Qwen3-8B
@ -14,6 +15,42 @@ license:
allows_local_ops: true
allows_fine_tune: true
notes: "Confirm exact card license at download time; Qwen3 line generally Apache-2.0."
profile:
summary: "Default mid-small open instruct for local ops, tools, and domain fine-tunes."
original_source: https://huggingface.co/Qwen/Qwen3-8B
use_cases:
- "Local NetKingdom / Coulomb assistant (chat, docs, runbooks)"
- "Tool-using agent loops on consumer GPU"
- "QLoRA / LoRA domain specialization base"
- "Multilingual (incl. DE/EN) drafting and summarization"
- "Offline code help when 14B+ is too heavy"
sweet_spots:
- "Best balance of quality vs VRAM in the ~8B open class for many 2026 stacks"
- "Instruction + tool-use oriented workflows"
- "Homelab fine-tune target (axis C)"
- "Runnable spine default when one model must wear many hats"
not_ideal_for:
- "Hardest SWE-bench-class multi-file engineering (use larger or closed frontier)"
- "Deep multi-step math/reason vs R1-class distill or full reasoners"
- "Embedding / retrieval (use BGE-M3 or nomic)"
capability_notes: >
Flagship small-mid dense open generalist in the Qwen3 line. Strong multilingual
and instruct behavior for its size; primary R-tier workhorse for the lab. Prefer
Instruct sibling on the card if separate repo exists at pin time.
swot:
strengths:
- "High capability density at 8B; Apache-friendly licensing typical"
- "Good multilingual + tool/instruct posture for local agents"
- "Excellent FT base for domain specialization"
weaknesses:
- "Still far from frontier closed models on hard agentic coding"
- "8B ceiling on long-horizon planning and rare knowledge"
opportunities:
- "Domain LoRAs (security, ops, railiance) on NAS-held base"
- "Pair with BGE-M3 RAG for grounded NetKingdom answers"
threats:
- "Rapid supersession by next Qwen/peer 8–14B release"
- "Quant quality variance across third-party GGUF repacks"
size:
total_bytes: 0
total_human: "~16 GB fp16 / ~5 GB Q4 (estimate)"
@ -31,8 +68,8 @@ collection:
storage_path: ""
brief_refs:
- research/2026-07-24-baseline-field-survey.md
reason: "P0 spine — best default open general/tool model for local ops and QLoRA domain specialization."
tags: [instruct, text, qwen3, ft-base]
reason: "P0/R spine — best default open general/tool model for local ops and QLoRA domain specialization."
tags: [tier-r, instruct, text, qwen3, ft-base]
companions: []
notes: "Prefer Instruct variant on card if separate repo; pin commit SHA at collection."
history:
@ -40,3 +77,7 @@ history:
event: nominated
by: baseline-survey
detail: "P0 recommendation from initial deep research."
- at: "2026-07-24"
event: profile_swot_added
by: grok
detail: "schema 0.2 profile + SWOT."

50
inventory/catalog/deepseek-ai__DeepSeek-R1-Distill-Qwen-14B__candidate.yaml
View file

@ -7,6 +7,7 @@ source:
url: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B
revision: main
model_card_url: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B
project_url: https://github.com/deepseek-ai/DeepSeek-R1
license:
spdx: MIT
url: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B
@ -14,6 +15,42 @@ license:
allows_local_ops: true
allows_fine_tune: true
notes: "R1 distill series MIT — confirm card at pin time."
profile:
summary: "Local reasoning-specialist distill — hard tasks without full R1 MoE."
original_source: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B
use_cases:
- "Offline multi-step reasoning (math, logic, planning sketches)"
- "Harder agent tool-selection and self-checks"
- "Local alternative when DeepSeek API is unavailable"
- "Eval harness baseline for 'reason' quality on lab tasks"
- "QLoRA experiments on reason-style behavior"
sweet_spots:
- "Reasoning quality above plain 14B chat models at similar size"
- "Single enthusiast GPU (Q4) without multi-node MoE"
- "MIT-friendly open reason lineage"
not_ideal_for:
- "Cheapest always-on chat (use 3B/8B)"
- "Full parity with frontier closed reasoners or full R1"
- "Pure embedding"
capability_notes: >
Distilled from DeepSeek-R1 into a Qwen-14B-class dense student. Captures much
of the reasoner value proposition for homelab VRAM budgets. Primary R-tier
reason model; full R1 remains strategic S-tier.
swot:
strengths:
- "Strong open reasoner at single-GPU scale"
- "MIT distill story; clear provenance from R1 line"
- "Complements general 8B chat without replacing it"
weaknesses:
- "Heavier and slower than 8B for routine chat"
- "Distill ≠ full R1; still loses on hardest problems"
- "May overthink simple tasks if not prompted carefully"
opportunities:
- "Route only hard queries here; keep 8B as default"
- "Domain FT for structured ops decision checklists"
threats:
- "Newer distill or mid-size reasoners may obsolete this checkpoint"
- "VRAM growth of defaults may push lab to 32B distill instead"
size:
total_bytes: 0
total_human: "~28 GB fp16 / ~9 GB Q4 (estimate)"
@ -31,12 +68,17 @@ collection:
storage_path: ""
brief_refs:
- research/2026-07-24-baseline-field-survey.md
reason: "P0 local reasoning without full R1 MoE — agent/tool loops and harder offline tasks."
tags: [reasoning, distill, deepseek]
companions: []
notes: "If disk/VRAM constrained, substitute DeepSeek-R1-Distill-Qwen-8B as P0 alternate."
reason: "P0/R local reasoning without full R1 MoE — agent/tool loops and harder offline tasks."
tags: [tier-r, reasoning, distill, deepseek]
companions:
- deepseek-ai__DeepSeek-R1__strategic
notes: "If disk/VRAM constrained, substitute DeepSeek-R1-Distill-Qwen-8B as R alternate."
history:
- at: "2026-07-24"
event: nominated
by: baseline-survey
detail: "P0 recommendation from initial deep research."
- at: "2026-07-24"
event: profile_swot_added
by: grok
detail: "schema 0.2 profile + SWOT."

48
inventory/catalog/deepseek-ai__DeepSeek-R1-Distill-Qwen-32B__candidate.yaml
View file

@ -7,13 +7,47 @@ source:
url: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B
revision: main
model_card_url: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B
project_url: https://github.com/deepseek-ai/DeepSeek-R1
license:
spdx: MIT
url: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B
allows_offline_retention: true
allows_local_ops: true
allows_fine_tune: true
notes: ""
notes: "R1 distill series MIT — confirm card at pin time."
profile:
summary: "Larger R1 distill for stronger local reason when 14B is not enough."
original_source: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B
use_cases:
- "Hard offline reasoning beyond 14B distill quality"
- "Local eval ceiling before considering full R1 MoE"
- "Heavier agent planning loops on multi-GPU / high-VRAM hosts"
- "FT experiments on reason-style traces at 32B"
sweet_spots:
- "Best dense open reason step between 14B distill and full R1"
- "Still denser/portable than 600B-class MoE"
- "MIT-friendly R1 lineage"
not_ideal_for:
- "Default always-on chat (too heavy)"
- "Hosts without ~20+ GB VRAM for Q4"
- "When 14B distill already saturates task quality"
capability_notes: >
W-tier optional. Approve only with VRAM + soft-quota headroom after R spine
and primary S pulls. Do not confuse with full DeepSeek-R1 MoE (separate S entry).
swot:
strengths:
- "Material reason quality jump over 14B distill"
- "Dense, so simpler serving than full MoE"
- "Clear provenance in R1 distill family"
weaknesses:
- "VRAM and latency cost; poor default chat model"
- "Still not full R1; diminishing returns vs S2 full weights"
opportunities:
- "Route only hardest local tasks here"
- "Bridge until multi-GPU can host full R1"
threats:
- "Quota competition with S1/S4 strategic weights"
- "Newer mid-large open reasoners may leapfrog"
size:
total_bytes: 0
total_human: "~65 GB fp16 / ~20 GB Q4 (estimate)"
@ -31,12 +65,18 @@ collection:
storage_path: ""
brief_refs:
- research/2026-07-24-baseline-field-survey.md
reason: "P1 stronger local reasoner — approve only with VRAM + quota headroom."
tags: [reasoning, distill, deepseek]
companions: []
reason: "P1/W stronger local reasoner — approve only with VRAM + quota headroom."
tags: [tier-w, reasoning, distill, deepseek]
companions:
- deepseek-ai__DeepSeek-R1-Distill-Qwen-14B__candidate
- deepseek-ai__DeepSeek-R1__strategic
notes: "Do not collect full DeepSeek-V3/R1 MoE under this id."
history:
- at: "2026-07-24"
event: nominated
by: baseline-survey
detail: "P1 recommendation from initial deep research."
- at: "2026-07-28"
event: profile_swot_added
by: grok
detail: "schema 0.2 profile + SWOT."

84
inventory/catalog/deepseek-ai__DeepSeek-R1__strategic.yaml Normal file
View file

@ -0,0 +1,84 @@
id: deepseek-ai__DeepSeek-R1__strategic
status: candidate
name: DeepSeek-R1
org: deepseek-ai
source:
kind: huggingface
url: https://huggingface.co/deepseek-ai/DeepSeek-R1
revision: main
model_card_url: https://huggingface.co/deepseek-ai/DeepSeek-R1
project_url: https://github.com/deepseek-ai/DeepSeek-R1
license:
spdx: MIT
url: https://huggingface.co/deepseek-ai/DeepSeek-R1
allows_offline_retention: true
allows_local_ops: true
allows_fine_tune: true
notes: "MIT for R1 series per public release notes; re-confirm card at pin."
profile:
summary: "Strategic full open reasoner — best-known open R1-class weights for future multi-GPU."
original_source: https://huggingface.co/deepseek-ai/DeepSeek-R1
use_cases:
- "Offline retention of top open reasoner lineage"
- "Future local hosted reason when multi-GPU is available"
- "Teacher / reference for distill and eval harnesses"
- "Hard planning and math when API reasoners are unavailable"
sweet_spots:
- "S2 pure-reason niche on the reserve"
- "Complements V3 general MoE rather than replacing the 14B distill spine"
- "MIT-friendly open reasoner identity"
not_ideal_for:
- "Current single-GPU daily reason (use R1-Distill-14B)"
- "Full-precision pull if soft quota already holds S1 giant"
- "Casual chat default"
capability_notes: >
Full DeepSeek-R1 (or open successor reasoner). Distills cover runnable reason;
this entry is strategic optionality. If quota forces a choice between full R1
and full V3, re-evaluate from latest brief which niche is the larger gap.
swot:
strengths:
- "Reference open reasoner class for the era"
- "Distill ecosystem already proves value at smaller sizes"
- "Open weights for offline continuity"
weaknesses:
- "Enormous resource cost to serve"
- "Overlaps capability budget with S1 MoE on a small disk"
- "Version supersession risk"
opportunities:
- "Hold compressed if needed; keep identity for later upgrade"
- "Use as teacher target for lab distill experiments"
threats:
- "Quota: competing with S1 for hundreds of GiB"
- "Newer open reasoners may displace this checkpoint"
size:
total_bytes: 0
total_human: "large — similar class to V3 full; prefer compressed if needed for 850 GiB soft quota"
hardware_class:
min_vram_gb_q4: 0
min_vram_gb_fp16: 0
notes: "Beyond current lab run envelope. Distills (14B/32B) cover runnable reason; this is full strategic reasoner."
axes: [A, B, C]
priority: high
collection:
approved_by: ""
approved_at: null
downloaded_at: null
downloaded_by: ""
storage_path: ""
brief_refs:
- research/2026-07-24-nas-strategic-collection-plan.md
- docs/decisions/2026-07-24-nas-strategic-reserve.md
reason: "S2 strategic — best-known open reasoner lineage; optionality for future multi-GPU / hosted local reason."
tags: [tier-s, strategic, beyond-run-envelope, reasoning]
companions:
- deepseek-ai__DeepSeek-R1-Distill-Qwen-14B__candidate
notes: "If quota forces a choice between full R1 and full V3, prefer the stronger current open general (often V3/V4 line) unless pure reason is the gap; re-evaluate at pull time from latest brief."
history:
- at: "2026-07-24"
event: nominated
by: operator-policy
detail: "Strategic NAS reserve — unrunnable OK."
- at: "2026-07-28"
event: profile_swot_added
by: grok
detail: "schema 0.2 profile + SWOT."

83
inventory/catalog/deepseek-ai__DeepSeek-V3__strategic.yaml Normal file
View file

@ -0,0 +1,83 @@
id: deepseek-ai__DeepSeek-V3__strategic
status: candidate
name: DeepSeek-V3
org: deepseek-ai
source:
kind: huggingface
url: https://huggingface.co/deepseek-ai/DeepSeek-V3
revision: main
model_card_url: https://huggingface.co/deepseek-ai/DeepSeek-V3
project_url: https://github.com/deepseek-ai/DeepSeek-V3
license:
spdx: MIT
url: https://huggingface.co/deepseek-ai/DeepSeek-V3
allows_offline_retention: true
allows_local_ops: true
allows_fine_tune: true
notes: "Confirm active card license at pin; prefer latest V3.x/V4 open successor if it is the new SOTA open general."
profile:
summary: "Strategic open general MoE — among the most capable open chat/code weights."
original_source: https://huggingface.co/deepseek-ai/DeepSeek-V3
use_cases:
- "Long-term offline optionality if frontier API access is lost or restricted"
- "Future multi-GPU / training-facility inference or continued pretrain"
- "Capability benchmark reference against closed frontier"
- "Source of student weights / distill experiments (when license + hardware allow)"
sweet_spots:
- "S1 primary giant: best open general MoE class on the reserve"
- "Capability identity even when only compressed weights fit disk"
- "MIT-friendly open SOTA lineage (confirm at pin)"
not_ideal_for:
- "Current lab single-GPU daily ops (use R spine)"
- "Storing multiple near-duplicate full-precision giants on one disk"
- "Assuming immediate local serve without major hardware upgrade"
capability_notes: >
Flagship open general MoE (or successor) for the strategic tier. On ~1–2 TB
lab bulk storage, prefer official compressed distributions when full precision
would exhaust soft quota. Catalog identity stays the base model id + revision.
swot:
strengths:
- "Top-tier open general capability for the era"
- "Open weights enable true offline retention and future fine work"
- "Strong coding/agent relevance as hardware catches up"
weaknesses:
- "Huge disk and multi-GPU serve cost"
- "Rapid version churn (V3 → V3.x → V4)"
- "Quant choice materially affects quality"
opportunities:
- "Hold compressed primary; upgrade hardware later without re-scrape risk"
- "Anchor daily briefs against a fixed offline SOTA open baseline"
threats:
- "Soft quota: one full pull can block other S niches"
- "License or card terms change at successor releases — re-check at pin"
size:
total_bytes: 0
total_human: "hundreds of GiB (quant) to multi-hundred+ GiB; measure at pull — may need official compressed release"
hardware_class:
min_vram_gb_q4: 0
min_vram_gb_fp16: 0
notes: "Beyond current lab run envelope (T4+). Reserved for strategic capability, not current inference."
axes: [A, B, C]
priority: high
collection:
approved_by: ""
approved_at: null
downloaded_at: null
downloaded_by: ""
storage_path: ""
brief_refs:
- research/2026-07-24-nas-strategic-collection-plan.md
- docs/decisions/2026-07-24-nas-strategic-reserve.md
reason: "S1 strategic — among the most capable open general MoE weights; keep even if unrunnable today."
tags: [tier-s, strategic, beyond-run-envelope, moe, general]
companions: []
notes: "At download: pin exact revision; if full precision exceeds remaining soft quota, collect best official compressed distribution of the same model line."
history:
- at: "2026-07-24"
event: nominated
by: operator-policy
detail: "1TB NAS strategic reserve — capability-first, not run-envelope-gated."
- at: "2026-07-28"
event: profile_swot_added
by: grok
detail: "schema 0.2 profile + SWOT."

44
inventory/catalog/meta-llama__Llama-3.2-3B-Instruct__candidate.yaml
View file

@ -7,6 +7,7 @@ source:
url: https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct
revision: main
model_card_url: https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct
project_url: https://www.llama.com/
license:
spdx: custom
url: https://ai.meta.com/llama/license/
@ -14,6 +15,41 @@ license:
allows_local_ops: true
allows_fine_tune: true
notes: "Llama Community License — not MIT; review terms before commercial redistribution."
profile:
summary: "Tiny edge instruct for always-on and low-VRAM agent micro-services."
original_source: https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct
use_cases:
- "Always-on local helper with minimal power/VRAM"
- "Fast classification, routing, and short-form rewrites"
- "CPU or 4–8 GB GPU edge boxes"
- "Smoke-test harnesses before promoting to 8B+"
- "Embedded tooling demos and offline kiosks"
sweet_spots:
- "Latency and footprint over peak intelligence"
- "Huge Llama ecosystem (llama.cpp, Ollama, many templates)"
- "Cheap parallel micro-agents in a fleet"
not_ideal_for:
- "Serious coding or long documents"
- "Domain FT when quality matters (prefer 8B base)"
- "Standalone reasoning-heavy ops decisions"
capability_notes: >
Compact Meta instruct model. Best as a floor for local intelligence and
high-volume simple tasks, not as the lab's primary brain.
swot:
strengths:
- "Very small; runs almost anywhere in the lab"
- "Massive tooling and community template coverage"
- "Good enough for structured short outputs"
weaknesses:
- "Shallow capability; fails hard tasks silently if not constrained"
- "Custom Llama license (not pure OSS-permissive)"
- "HF gated access friction"
opportunities:
- "Fleet of specialized tiny adapters per workflow"
- "Guardrail / router model in front of larger backends"
threats:
- "Qwen/Gemma/Smol peers may outclass at same size"
- "License or regional policy changes for Llama family"
size:
total_bytes: 0
total_human: "~6 GB fp16 / ~2 GB Q4 (estimate)"
@ -31,8 +67,8 @@ collection:
storage_path: ""
brief_refs:
- research/2026-07-24-baseline-field-survey.md
reason: "P0 edge instruct — tiny, huge ecosystem, good CPU/GPU floor for agents."
tags: [instruct, edge, llama]
reason: "P0/R edge instruct — tiny, huge ecosystem, good CPU/GPU floor for agents."
tags: [tier-r, instruct, edge, llama]
companions: []
notes: "HF gated model — need accepted license on account before download."
history:
@ -40,3 +76,7 @@ history:
event: nominated
by: baseline-survey
detail: "P0 recommendation from initial deep research."
- at: "2026-07-24"
event: profile_swot_added
by: grok
detail: "schema 0.2 profile + SWOT."

84
inventory/catalog/meta-llama__Llama-3.3-70B-Instruct__strategic.yaml Normal file
View file

@ -0,0 +1,84 @@
id: meta-llama__Llama-3.3-70B-Instruct__strategic
status: candidate
name: Llama-3.3-70B-Instruct
org: meta-llama
source:
kind: huggingface
url: https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct
revision: main
model_card_url: https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct
project_url: https://www.llama.com/
license:
spdx: custom
url: https://ai.meta.com/llama/license/
allows_offline_retention: true
allows_local_ops: true
allows_fine_tune: true
notes: "Llama Community License — review before commercial redistribution. Prefer Llama 4 open text sibling if it is clearly stronger at pull time."
profile:
summary: "Strategic dense Llama 70B instruct — ecosystem-rich open reserve for multi-GPU."
original_source: https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct
use_cases:
- "Future multi-GPU general instruct with mature Llama tooling"
- "Portable dense alternative to giant MoE for lab upgrades"
- "Ecosystem baselines (vLLM, llama.cpp, eval harnesses expect Llama ids)"
- "Domain FT when Llama license fits the use"
sweet_spots:
- "S4 dense open with the broadest third-party tooling surface"
- "More portable serve path than 600B-class MoE"
- "Strong general instruct heritage in the 70B class"
not_ideal_for:
- "Current single-GPU daily ops (use 3B/8B R spine)"
- "Uses forbidden by Llama Community License terms"
- "Duplicate with Qwen-72B under tight quota — pick one first"
capability_notes: >
HF gated. Strategic dense pick. If Llama 4 open text weights supersede
clearly, nominate successor and mark this entry superseded. License is not
Apache — operator must re-read terms before redistribution.
swot:
strengths:
- "Huge ecosystem and ops familiarity"
- "Solid 70B dense instruct quality"
- "Easier multi-GPU dense story than MoE giants"
weaknesses:
- "Custom Llama license (not Apache/MIT)"
- "HF gating friction for automated pulls"
- "May lag peer open MoE on some benches"
opportunities:
- "First dense strategic if Qwen-72B is deferred"
- "Broad eval comparability with published Llama numbers"
threats:
- "Llama 4 open may obsolete 3.3 quickly"
- "Quota competition with Qwen-72B and S1 MoE"
size:
total_bytes: 0
total_human: "~140 GB fp16 / ~40 GB Q4 (estimate)"
hardware_class:
min_vram_gb_q4: 40
min_vram_gb_fp16: 140
notes: "T3+ to run well; still valuable reserve for future multi-GPU."
axes: [A, B, C]
priority: high
collection:
approved_by: ""
approved_at: null
downloaded_at: null
downloaded_by: ""
storage_path: ""
brief_refs:
- research/2026-07-24-nas-strategic-collection-plan.md
- docs/decisions/2026-07-24-nas-strategic-reserve.md
reason: "S4 strategic dense — strong ecosystem 70B open instruct; more portable than full MoE for a future lab upgrade."
tags: [tier-s, strategic, dense, instruct, llama]
companions:
- meta-llama__Llama-3.2-3B-Instruct__candidate
notes: "HF gated. If Llama 4 open weights supersede, nominate successor and supersede this entry."
history:
- at: "2026-07-24"
event: nominated
by: operator-policy
detail: "Strategic dense open for NAS capability reserve."
- at: "2026-07-28"
event: profile_swot_added
by: grok
detail: "schema 0.2 profile + SWOT."

45
inventory/catalog/nomic-ai__nomic-embed-text-v1.5__candidate.yaml
View file

@ -7,6 +7,7 @@ source:
url: https://huggingface.co/nomic-ai/nomic-embed-text-v1.5
revision: main
model_card_url: https://huggingface.co/nomic-ai/nomic-embed-text-v1.5
project_url: https://www.nomic.ai/
license:
spdx: Apache-2.0
url: https://huggingface.co/nomic-ai/nomic-embed-text-v1.5
@ -14,6 +15,39 @@ license:
allows_local_ops: true
allows_fine_tune: true
notes: ""
profile:
summary: "Lightweight open text embedder for A/B with BGE-M3 and long-context retrieval experiments."
original_source: https://huggingface.co/nomic-ai/nomic-embed-text-v1.5
use_cases:
- "Secondary / A/B embedding path vs BGE-M3"
- "English-heavy RAG with low resource cost"
- "Long-context embedding experiments (per card capabilities)"
- "Portable laptop offline search indexes"
sweet_spots:
- "Small footprint Apache embed"
- "Fast iteration when re-embedding corpora often"
- "Nomic tooling and Matryoshka-style dimension flexibility (per card)"
not_ideal_for:
- "Sole multilingual production embed without A/B (prefer BGE-M3 default)"
- "Generation tasks"
- "Code-only retrieval (prefer code embed if held)"
capability_notes: >
Compact Nomic text embedding model. Held as runnable dual-path to avoid
single-vendor embed lock-in and to re-eval retrieval quality cheaply.
Prefer current v2 text card at pull if clearly better.
swot:
strengths:
- "Tiny, Apache-2.0, easy to re-run indexes"
- "Diversifies embed dependency next to BGE-M3"
weaknesses:
- "May lose multilingual or domain A/B vs BGE-M3"
- "Version churn (v1.5 vs v2) needs careful pin"
opportunities:
- "Dimension-reduced storage for large corpora"
- "Hybrid ensembles (BGE + nomic signals)"
threats:
- "Becomes redundant if one embed wins all lab evals"
- "Upstream rename/deprecation of v1.5"
size:
total_bytes: 0
total_human: "<1 GB"
@ -31,12 +65,17 @@ collection:
storage_path: ""
brief_refs:
- research/2026-07-24-baseline-field-survey.md
reason: "P0 lightweight embed for A/B with BGE-M3; long-context text retrieval."
tags: [embedding, rag]
companions: []
reason: "P0/R lightweight embed for A/B with BGE-M3; long-context text retrieval."
tags: [tier-r, embedding, rag]
companions:
- BAAI__bge-m3__candidate
notes: "If v2 text is preferred at collection time, update id/url and supersede this candidate."
history:
- at: "2026-07-24"
event: nominated
by: baseline-survey
detail: "P0 recommendation from initial deep research."
- at: "2026-07-24"
event: profile_swot_added
by: grok
detail: "schema 0.2 profile + SWOT."