consult-nvidia-models

name

consult-nvidia-models

description

Multi-model fan-out across NVIDIA Integrate API (NIM). Lists all available text-gen models (DeepSeek v4 Pro, MiniMax, Kimi-K2.6, Nemotron, GPT-OSS, Z-AI GLM, Qwen, Llama-4, Mistral, etc.) and fans a single prompt to N of them in parallel, then aggregates replies. Use when the user says "/consult-nvidia-models", "/consult-nvidiamodels", "ask nvidia models", "nvidia multi-model", "nvidia fan-out", or wants triangulation across multiple NIM-hosted models. Subcommands - list, search, run, fan-out.

consult-nvidia-models

Fan a single prompt across multiple NVIDIA-hosted text-generation models in one call. Complements consult-nvidia-deepseek (single-model) by being multi-model-by-default.

When to use

• User says /consult-nvidia-models, /consult-nvidiamodels, "ask nvidia models", "nvidia fan-out", "nvidia second opinions"
• User names ≥2 NIM-hosted models: deepseek-v4-pro, minimax-m2.7, kimi-k2.6, nemotron-ultra-253b, gpt-oss-120b, glm-5.1, qwen3-coder-480b, llama-4-maverick, mistral-large-3-675b
• User wants a triangulated second-opinion call across families (e.g. "what do MiniMax + Kimi + GLM all say")
• User wants the list of available models before picking
• User wants to fan a quant/audit/strategy question to a panel of models

For single-model NIM calls, prefer consult-nvidia-deepseek — it's optimised for that path.

Prerequisites

Set the API key in env (never hardcode):

export NVIDIA_API_KEY='nvapi-...'   # preferred name
# or NGC_API_KEY (also accepted by tools/swarm/api_consult.py)

Verify the key works:

curl -s --max-time 10 "https://integrate.api.nvidia.com/v1/models" \
  -H "Authorization: Bearer $NVIDIA_API_KEY" \
  | python3 -c "import sys,json; print('OK,', len(json.load(sys.stdin).get('data',[])), 'models')"

🔴 MANDATORY: Leakage-Context Block

Every prompt sent to NIM models for asset-class / edge / pick-performance questions MUST include a leakage-context block listing known data-quality incidents that intersect the prompt's domain. Multi-model consensus is worthless when all models see the same corrupted inputs.

Why this exists (case study, 2026-05-25)

5/5 NIM models (Kimi K2.6, GPT-OSS-120B, GLM-5.1, Nemotron Super 49B, Mistral Nemotron) converged on "COMMODITY is the system's #1 alpha, allocate 20-30%" — because the prompt fed them pre-dedup numbers (PF=4.31) without mentioning H-101. In parallel, a 3-engine Codex/Gemini/Grok panel that WAS shown the leakage evidence correctly classified the same cell as DATA_QUALITY_LEAKAGE at ~90% confidence. The difference was the prompt, not the model quality. Reports: reports/2026-05-25_multi_ai_panel_meta_review.md + reports/2026-05-25_commodity_cot_edge_triangulation.md.

Template — paste at the TOP of any analysis prompt

LEAKAGE CONTEXT — Read before analyzing:

The following known data-quality incidents affect the dataset you are about
to analyze. Your first task is to determine whether ANY of these incidents
invalidate the numbers presented. If so, flag it explicitly and adjust your
analysis accordingly.

Known incidents (see reports/hypothesis_registry.json for full state):
 - H-101 (COMMODITY): COT publication-lag look-ahead. Pre-dedup commodity
   numbers show PF~4.3/WR~58%; post-policy ground truth is PF~0.31/WR~11%.
   The CFTC COT release-lag (T+3) was not respected. Any commodity analysis
   citing PF>1.0 on multi_asset_copytrader or multi_asset_cot is
   contaminated unless proven otherwise.
 - M-095 (COMMODITY): CT=F cotton over-emission. 230 closed picks come from
   16 unique signal dates (14× re-emission ratio). One profitable trade
   counted ~14×, inflating PF.
 - M-069 (NET): slippage-adjusted re-judgment reclassifies marginal wins
   as losses. Gross-PF and Net-PF can diverge by 2-3×.
 - QUALITY_BUCKETS (ALL): the `quality_bucket` field
   (profitable_tp / moderate_confidence / etc.) in
   performance_report_*.json is POST-HOC and circular. `profitable_tp` =
   "picks that hit TP." Do NOT treat the bucket as a pre-trade filter.

Be skeptical:
 - If the data suggests one symbol/source dominates, flag concentration risk.
 - If you see PF>3.0 on a single asset class, check for look-ahead or
   survivorship bias before recommending allocation.

When to include the block

• ALWAYS for asset-class performance questions citing dashboard numbers
• ALWAYS for COMMODITY questions (H-101 / M-095 must be stated)
• ALWAYS when referencing quality_bucket data in performance_report_*.json
• SKIP only for purely theoretical / no-dataset prompts

Post-fan-out cross-check

After collecting model replies, flag any model that endorses a known-bad signal without addressing the cited leakage evidence. Document the failure mode; it's diagnostic of the model's grounding behavior.

Subcommands

list — show all text-gen models on NIM

curl -s --max-time 30 "https://integrate.api.nvidia.com/v1/models" \
  -H "Authorization: Bearer $NVIDIA_API_KEY" \
  | python3 -c "import sys,json; [print(m['id']) for m in json.load(sys.stdin).get('data',[])]"

search <keyword> — filter the list

KW="${KW:-glm}"
curl -s --max-time 30 "https://integrate.api.nvidia.com/v1/models" \
  -H "Authorization: Bearer $NVIDIA_API_KEY" \
  | python3 -c "
import sys, json, os
kw = os.environ['KW'].lower()
for m in json.load(sys.stdin).get('data', []):
    if kw in m['id'].lower(): print(m['id'])
"

Useful keywords: deepseek, minimax, kimi, nemotron, gpt-oss, glm, qwen, llama-4, mistral, phi.

run <model> <prompt> — single-model call

Use consult-nvidia-deepseek for this. Or directly:

MODEL="${MODEL:-deepseek-ai/deepseek-v4-pro}"
PROMPT="${PROMPT:-Give me a one-sentence answer.}"
curl -s --max-time 90 "https://integrate.api.nvidia.com/v1/chat/completions" \
  -H "Authorization: Bearer $NVIDIA_API_KEY" \
  -H "Content-Type: application/json" \
  -d "$(python3 -c "
import json, os
print(json.dumps({
  'model': os.environ['MODEL'],
  'messages': [{'role': 'user', 'content': os.environ['PROMPT']}],
  'max_tokens': 1500,
  'temperature': 0.3,
}))")" \
  | python3 -c "import sys,json; print(json.load(sys.stdin)['choices'][0]['message']['content'])"

fan-out — fan one prompt to N models, sequential calls (the headline command)

Save prompt to a file (avoids quoting hell), pick a model panel, fan out:

# 1. Pick the panel — diversify families
PANEL=(
  "deepseek-ai/deepseek-v4-pro"
  "minimaxai/minimax-m2.7"
  "moonshotai/kimi-k2.6"
  "nvidia/llama-3.1-nemotron-ultra-253b-v1"
  "openai/gpt-oss-120b"
  "z-ai/glm-5.1"
)

# 2. Write the prompt to a file
cat > /tmp/nim_prompt.txt <<'EOF'
Your prompt here. Keep it self-contained; the models have no tool/file access.
Tell them: be quantitative, cite the numbers in the prompt verbatim, do not invent.
EOF

# 3. Fan out (sequential — parallelism via & is fine but rate-limits bite)
mkdir -p /tmp/nim_fanout
for MODEL in "${PANEL[@]}"; do
  SLUG=$(echo "$MODEL" | tr '/.-' '___')
  PROMPT="$(cat /tmp/nim_prompt.txt)" python3 -c "
import json, os, sys, subprocess, time
model = '$MODEL'
payload = {
  'model': model,
  'messages': [{'role': 'user', 'content': os.environ['PROMPT']}],
  'max_tokens': 1500,
  'temperature': 0.3,
}
r = subprocess.run(['curl','-s','--max-time','90',
  'https://integrate.api.nvidia.com/v1/chat/completions',
  '-H', f'Authorization: Bearer {os.environ[\"NVIDIA_API_KEY\"]}',
  '-H', 'Content-Type: application/json',
  '-d', json.dumps(payload)], capture_output=True, text=True, timeout=95)
try:
    d = json.loads(r.stdout)
    out = d['choices'][0]['message']['content']
except Exception as exc:
    out = f'ERROR: {exc}; raw={r.stdout[:400]}'
print(out)
" > "/tmp/nim_fanout/$SLUG.md" 2>&1
  echo "  -> /tmp/nim_fanout/$SLUG.md ($(wc -c < /tmp/nim_fanout/$SLUG.md) bytes)"
  sleep 0.5  # polite spacing
done

# 4. Aggregate
for f in /tmp/nim_fanout/*.md; do
  echo "=== $(basename "$f" .md) ==="
  head -20 "$f"
  echo
done

Curated panels (for common use cases)

# Wide diversity (use as default for triangulation)
PANEL_WIDE=(
  "deepseek-ai/deepseek-v4-pro"
  "minimaxai/minimax-m2.7"
  "moonshotai/kimi-k2.6"
  "openai/gpt-oss-120b"
  "z-ai/glm-5.1"
)

# Reasoning-heavy
PANEL_REASON=(
  "nvidia/llama-3.1-nemotron-ultra-253b-v1"
  "nvidia/nemotron-3-super-120b-a12b"
  "qwen/qwen3.5-397b-a17b"
  "mistralai/mistral-large-3-675b-instruct-2512"
)

# Coding-heavy
PANEL_CODE=(
  "qwen/qwen3-coder-480b-a35b-instruct"
  "mistralai/codestral-22b-instruct-v0.1"
  "deepseek-ai/deepseek-coder-6.7b-instruct"
  "openai/gpt-oss-120b"
)

Pitfalls

• NVIDIA NIM keys hang silently on some models if you exceed per-key rate. If a model takes >60s with no chunks, kill it and try a sibling.
• temperature defaults differ by model; pin to 0.3 for analysis/triangulation, 0.7+ for brainstorm.
• max_tokens: NIM caps differ. 1500 is safe; 4096 works for most reasoning models.
• Streaming: omit stream: true for fan-out — non-stream is easier to capture per-model.
• Output volume: each reply ~1.5-4 KB. Fan-out of 5 models = ~15 KB total — fine for a session, but pipe to files instead of stdout if you're going wider than 10 models.
• Cross-model disagreement is the whole point. Look at the distribution of answers, not the average. The 2026-05-25 commodity-edge triangulation (3 engines all flagged DATA_QUALITY_LEAKAGE at ~90% confidence) is the gold-standard pattern.

When NOT to use

• A question that needs file/web access — NIM models are text-in/text-out only. Feed them the JSON/file content verbatim.
• A question one model would answer just as well — use consult-nvidia-deepseek to save tokens.
• Operational-safety questions (e.g. "should I run this SQL on prod?") — escalate to a human, not a multi-model vote.

Related

• consult-nvidia-deepseek — single-model NIM consult (this skill's sibling)
• consult-cloudflare-models — same shape but Cloudflare Workers AI provider
• swarm-run — local swarm dispatcher that already supports NIM/Cloudflare as engines
• swarm-second-opinion — 3-engine quick consensus across mixed providers
• tools/swarm/api_consult.py — backend the existing consult-* skills call