meituan's stealth model exposed: longcat vs qwen in 3 coding tasks

owl alpha hits #4 on open router by token spend per week – we tested why

owl alpha is a free model by "stealth" – a placeholder provider name used for pre-release, experimental models from unnamed labs or tech companies provided for free testing before official release

we asked owl alpha what it is. it said: "i'm longcat, an ai language model created by meituan."

@meituan is china's dominant local services super-app:

- food delivery (65%+ market share)
- hotel bookings, grocery, bike-sharing
- 770m+ annual users, listed in hong kong since 2018
- expanding internationally as "keeta" (hong kong, saudi arabia, uae, brazil)

meituan went deep on ai – releasing their own llm family (longcat) and an ai agent called xiaomei

so we put @Meituan_LongCat against @Alibaba_Qwen 3.7 max – two chinese internet giants, one ai fight. three tasks:

1. marshmallow on bonfire – a self-contained html pixel-art animation, 8px grid, animated flame and a marshmallow that droops, browns, and drips over a 4s loop

2. car over hills – scrolling pixel-art landscape with parallax layers, a bobbing car, drifting clouds, ~60fps

3. orcs attacking a fortress – marching orc sprites, a battering ram, arcing arrows, defenders on the battlements

two things stood out:

• first – token spend. longcat performs at roughly the same level as qwen, but uses 2–3x more tokens per task. the model requires more corrections and follow-up edits to get the output right. that matters a lot: when stealth mode ends and meituan starts charging for api access, the cost per task will be significantly higher – even if the per-token price is low

• second – speed. longcat runs at 40–60 tok/s. for a cloud model, that's not fast

still, the conclusion is clear: another chinese lab is about to enter the market with a competitive model – and, as expected, at lower prices