The 8B dense entry point of the Griffin lineup. Lite carries the full Aegis attention stack and the same structured trace contract as its larger siblings, sized to fit the PR-time and CI deep-scan budgets where every second of latency is visible to a developer.
Sub-1.5s round-trips fit inside the developer's pull-request inner loop. Lite carries enough capacity to explain one taint path and propose a sanitiser-aware fix without breaking flow.
Designed for the CI step that fans out across thousands of candidate findings. Throughput per dollar is the lowest of the Griffin family, with a tight latency budget that keeps PR checks under typical job timeouts.
Generates a one- or two-hunk patch with a cited sanitiser, plus a short rationale. Good for the common case where the bug is local and the fix is mechanical — bumps, replacements, allow-list tightenings.
Lite is the cheapest variant that still emits a structured trace. For single-path reachability questions it produces hypothesis, cited hops, disproof attempt, and patch in one pass without escalating.
Honest limits. The router escalates to heavier variants when the path needs more reasoning depth than 8B can carry.
32k context tops out around a few hundred files of call-graph metadata. For audits that need to fold in the entire transitive closure, escalate to Griffin M or L.
Long disproof chains across package boundaries need the deeper reasoning budget of Griffin L or Zero. Lite will produce a hypothesis but the disproof pass is unreliable past three hops.
Lite is sized for shared and dedicated cloud tiers. Air-gapped and sovereign deployments with the heaviest reasoning needs route to Griffin Zero on a sovereign cluster.
| Parameters | 8B (dense) |
| Context window | 32k tokens |
| p95 latency | ~1.2s end-to-end |
| Active params per token | 8B (dense, no MoE) |
| Quantisation | FP16 default, INT8 available |
| Deployment tier(s) | Shared cloud, Dedicated cluster |
| Minimum GPU | 1x A100 40GB |
| Recommended GPU | 1x A100 80GB or H100 |
| Memory footprint | ~24 GB at FP16 |
| Inference cost relative tier | Lowest of the Griffin variants |
| Eval - exploit-hypothesis accuracy | 68% |
| Eval - adversarial prompt resistance | 94% |
| Eval - security-Q&A hallucination rate | 1.4% |
A finding comes in, Lite produces hypothesis, cites the path, attempts a disproof against the project's sanitiser config, and emits a patch hunk. The whole exchange fits inside a normal PR comment cycle.
Renders the structured reasoning trace into a reviewer-friendly comment with the cited hops, the refutation that failed, and the proposed fix. The trace ships with the comment so reviewers can audit it.
After a CLI deep-scan, Lite condenses the per-finding traces into a top-priorities summary with rationales. Cheap enough to run across thousands of candidate findings in a single CI job.
When the on-device student model's confidence falls below threshold, the IDE routes that single finding to Lite for a deeper reasoning pass without leaving the inner loop.
A single-finding reasoning call with the latency budget pinned to the PR-time tier.
curl -X POST https://api.safeguard.sh/v1/reason \
-H "Authorization: Bearer $TOKEN" \
-H "Content-Type: application/json" \
-d '{
"variant": "griffin-lite",
"input": {
"finding_id": "fnd_8f2c19a4",
"mode": "single-finding",
"include_trace": true,
"include_patch": true
},
"constraints": {
"max_latency_ms": 1500,
"context_budget_tokens": 32000
}
}'Lite handles the low-complexity, single-path band of the Eagle triage score. That covers findings where the call graph has shallow depth, the sink severity is moderate, sanitiser ambiguity is low, and no cross-package edges are involved. Inside this band, Lite produces a structured trace at roughly a tenth of the L-tier compute cost.
Cheap enough to run on every pull request, deep enough to emit a structured trace with the finding.