671B parameters via 8-expert mixture-of-experts, ~37B active per token, 256k usable context. The deepest reasoning the lineup ships — sized for sovereign and air-gapped deployments, cross-package chains beyond twelve hops, and the agentic disclosure workflows that the other variants cannot sustain.
The deepest reasoning Safeguard ships, running inside the customer's own perimeter with no egress. Built for the longest reasoning budgets — quarterly portfolio audits, regulator-grade findings, classified-environment review.
Where Griffin L tops out, Zero keeps reasoning. 256k usable context plus the MoE expert routing carry chains through five or more package boundaries without dropping the cited path.
Agentic workflow: propose the upstream patch, run it through the maintainer's project test suite, and draft the disclosure thread. Zero is the only variant with the reasoning budget to sustain the full chain autonomously.
Multi-step autonomous reasoning over tools, retrieval, and the disproof head — workflows that need to plan, execute, observe, and replan five or six times. Other variants lose the thread; Zero does not.
Zero is opt-in by deployment tier. The auto-router will not silently send shared-cloud traffic to Zero — too expensive, too slow for what most workloads need.
A ~12s p95 plus the cost profile is wrong for inline PR review. Use Griffin Lite for IDE-side cloud-burst reasoning, and Griffin S for PR-level checks where the budget is seconds, not a session.
If a deployment gate must finish in under ten seconds, Zero is the wrong shape. Route the gate to Griffin M or L — they will carry the typical taint chain at a fraction of the wall-clock cost.
Zero is the highest-cost tier in the lineup. The auto-router avoids it by default on the shared-cloud tier — Zero is opt-in by deployment tier, not the cheapest variant that could carry a chain.
| Parameters | 671B total (sparse mixture-of-experts) |
| Context window | 256k usable (retrieval gates over windowed attention) |
| p95 latency | ~12s end-to-end |
| Active params per token | ~37B (~5.5% of total) |
| Quantisation | FP16 default · FP8 for sovereign deployments |
| Deployment tier(s) | Sovereign · Air-gapped · Dedicated cluster (large) |
| Minimum GPU | 8x H100 80GB |
| Recommended GPU | 11x H100 multi-AZ (Growth) up to 22x H100 multi-AZ (Mature) |
| Memory footprint | ~340 GB at FP8 · ~680 GB at FP16 |
| Inference cost relative tier | Highest |
| Eval — exploit-hypothesis accuracy | 88% |
| Eval — adversarial prompt resistance | 99% |
| Eval — security-Q&A hallucination rate | 0.3% |
| Eval — top-5 candidate-path retention vs known CVE ground truth | 97% (+12 precision points over Griffin L on cross-package taint paths) |
Run inside the customer's own region, account, and VPC, with no traffic leaving the perimeter. Zero is the reasoning tier sized for the regulatory and operational constraints of sovereign infrastructure.
Bring-your-own-cluster, no network egress, no outbound telemetry. Zero ships as a container image plus weights; the eval gate and disproof head run entirely on customer hardware.
End-to-end agentic flow: confirm the chain, draft the upstream patch, run the maintainer's test suite locally, and prepare the disclosure thread for a human to approve. Zero owns each step.
Teacher model for the heaviest Lino distillation experiments. The Lino-class students that ship to IDEs inherit their reasoning priors from Zero's traces under sustained agentic load.
Sovereign and air-gapped tiers can request Zero directly. The agentic block opts the call into upstream patch drafting, project-test execution, and disclosure preparation.
$ curl https://api.safeguard.sh/v1/reason \
-H 'Authorization: Bearer $SAFEGUARD_API_KEY' \
-H 'Content-Type: application/json' \
-d '{
"model": "griffin-zero",
"mode": "sovereign-audit",
"project": "proj_a91f...",
"deployment_tier": "sovereign",
"candidate": {
"cwe": "CWE-918",
"entry": "POST /api/webhook/dispatch",
"sink": "HttpClient.send(URI)"
},
"context_budget": 262144,
"emit_trace": true,
"run_disproof": true,
"agentic": {
"draft_upstream_patch": true,
"run_project_tests": true,
"draft_disclosure": true
}
}'
{
"model": "griffin-zero",
"finding_id": "find_8801",
"verdict": "exploitable",
"trace": {
"hypothesis": "CWE-918 via webhook URL allow-list bypass",
"cited_path_hops": 14,
"package_boundaries": 5,
"disproof_result": "refutation_failed",
"upstream_patch_applied": true,
"project_tests_passed": "412/412",
"disclosure_draft_ready": true
},
"experts_activated": ["taint", "http", "sanitiser-ambiguity"],
"latency_ms": 11904,
"router_score": 0.94
}Zero is opt-in by deployment tier. The default auto-router does not escalate to Zero on the shared-cloud tier — the cost and latency are wrong for typical production traffic. Zero is selected only when one of three conditions holds.
model: "griffin-zero" to force the variant.Sovereign-grade reasoning on hardware you control. Book a session to size the cluster and walk the agentic disclosure pipeline end-to-end.