Laya: a free, local alternative to Jev — and it can even play Doom(ish)
Medium post by Christian Graham (Friends of the Earth, Experiments team co-lead), published
2026-09-18. Held in raw/ since before this ingest, unfiled until now — folded in here because it
is a second, independent account of laya on the exact day the model shipped.
What it is
Not a benchmark. Graham read Laya's own comparison table against Jev ("free, local, and — on paper — about 10x faster and slightly more accurate"), then spent an afternoon pointing it at something the table doesn't test: playing Doom, where being wrong gets the agent killed rather than mis-scored.
The build, roughly: a single focused noul (yes/no) question — "should I shoot?" — worked where
folding shoot into a large choice list didn't; the model would never fire when shooting had to
compete against everything else. Hand-written guard rules sit on top for cases the model gets wrong
in ways that would otherwise go unnoticed (firing with no ammo, firing at nothing lined up), each one
logged rather than silently overridden. Exploration — clearing a room and then finding the way out —
is the part that doesn't work yet: the agent still circles a cleared room instead of moving on, and
the fixes tried (measuring the actual per-move turn angle, systematically trying doors and walls, a
last-resort wall-following rule) improved survival and exploration without solving it.
What it's evidence of, and what it isn't
One person, one weekend project, no measurement beyond "it didn't get instantly stuck." Graham's own framing states the limits plainly: Doom is "messy and only ever half-visible," which is the point — it stresses the model somewhere Laya's own benchmark table can't, and the honest result is "fights, survives, explores properly now... has yet to find its way out." That is a single practitioner's qualitative account, not a scored comparison, and it corroborates Laya's model card only on the cheapest claim: that the thing runs, locally, on ordinary hardware ("runs fine on my own Mac"), and is fast enough to drive a real-time loop.
The framing point worth keeping past the game — a calibrated, non-autoregressive model is a fit "anywhere you need a fast, cheap, trustworthy yes/no or pick-one answer rather than generated text": support-ticket and email routing, content flagging with a confidence score, deciding what needs a human, high-volume classification where a per-item LLM call is too slow, and anywhere the data can't leave the building. That list is the same shape as the escalate/act primitive on Laya's own model card, restated from a user's-eye view rather than a vendor's.
Tier T4 — a named practitioner arguing from one weekend's experience, with no data attached; see
../QUALITY.md → the practitioner argument. It isn't filler and it isn't a vendor speaking, it's
simply unevidenced by the ladder's own definition.
Related
laya · system-one-models-and-jev · convai-innovations · fine-tuning