Skip to content

RTX 5090 private candidate families and Ignis runtime trials

Current status and review date

Decision snapshot

  • Product role: bounded coding and synthetic capability research.
  • Best-tested settings: private candidate/Ignis 0.8.0 with 4,096 thinking budget; QAT/NInfer no-speculation; coding-focused/NInfer 128K text profile.
  • Measured hardware: one RTX 5090, Windows/Docker Linux VM.
  • Evidence: quality, functional and short-output capacity diagnostics; selected profiles each passed 123/123 coding checks.
  • Decision: challenger evidence, no-promotion.
  • Important limitation: private identities/text withheld; seven task types repeated three times, no real repository or equivalent-main qualification.
  • Review dates: evidence and dossier reviewed 2026-11-03.

Campaign finding · Ignis improvement note · Configuration description.

The coding candidate passed the fixture with lower observed task latency than the restored reference, but its 128K text profile is not the 160K vision profile. Ignis's budget follow-up recovered correctness at a substantial latency cost. QAT's text survivor did not qualify its failed vision sibling. Outcome: retain configuration-specific research leads; no replacement decision.

Immutable identity

Model/variant, weight/draft/projector, served identity and image fingerprints are withheld by the private-model boundary. Ignis 0.8.0 is a public runtime identity, not a model identity. Original reference capture was incomplete. Sanitized native artifacts do not grant public immutable-identity qualification or independent reproduction. See redactions.

Tested hardware and topology

One RTX 5090, 32 GB class; one active candidate at a time; direct online endpoints. No other GPU was measured. No routed-client or deployment claim follows.

Engine, quantization, KV, context, and concurrency recipe

The sanitized configuration records distinct 32K/128K/160K/256K profiles, quantization and available KV descriptions, host caps, thinking policy and configured lanes. It is not an executable recipe: exact identity is intentionally private. Four Ignis lanes are not four measured full-context requests.

Evidence by measurement class

Status: completed bounded quality/functional/short-output diagnostics; full serving comparison incomplete.

Measured: selected candidate profiles 123/123 coding; restored reference 114/123. Ignis large-profile retrieval 45/45 through 245,646 input tokens, tools 30/30, required image facts 18/18. QAT text retrieval 45/45 and tools 30/30. The coding candidate's earlier raw retrieval 225/225 remains inside a failed outer job, separate from its fresh coding recheck.

Limits: cross-profile policies and capabilities differ. Retrieval is synthetic; image markers do not score incidental hallucinations; tools are short-message fixtures; capacity uses 16-word C1 outputs. No broad ranking or real repository score.

Evidence: complete coding rows, summary, native index.

Decision and promotion state

Selected settings are challenger research leads, not current deployment assignments. The larger reasoning budget warrants held-out validation; QAT speculative failures and vision load failure remain rejected configurations. No serving route or model assignment changed through this publication.

Failures and gotchas

  • Reasoning: smaller budget broke state invariants; larger budget is not a universal optimum.
  • Vision: Ignis captions added unsupported interpretations; metadata reported text.
  • Memory: Ignis reached its 16GiB host cap without OOM; four lanes do not establish full-window capacity.
  • Evidence: failed outer jobs stay failed; historical identity gaps and public redactions remain visible.
  • Missing gates: repository tasks, routed clients, video, endurance and full-context concurrency.

Dated run history