Creator @teortaxesTex argues Western labs struggle with small LLMs because their scaling laws are optimized for trillion-parameter architectures

Not really They generate far more revenue on the flagships Training small models and serving them for cheap will hurt their topline

DSV4-Flash is loved. Gemini 3 Flash is loved. The first Haiku was deemed blessed. o4-mini was great. Yes they don't get as much mythmaking as flagships, but… I think the issue with the Western frontier is that they just *ain't the best* at small. Their scaling laws are for 1T+.

@zephyr_z9 I don't know if "serving flagships = more revenue" is a law. The bottleneck is supply. If they could sell 2x more Sonnet tokens than Opus tokens, but Sonnet were 3x cheaper to serve, by this logic they'd go all in on Sonnet. So why isn't this the case? mostly cuz Sonnet is bad

@teortaxesTex I think labs just don't care about small models

they could train much stronger Sonnet or Mini models but simply don't care enough

it doesn't benefit them in any way

Anthropic in particular. As they get better at Opus and now Fable and Mythos, they clearly abandon the mid- and low range. It didn't have to be like this, but they have a limited bandwidth.

@teortaxesTex I don't think Sonnet is 3x cheaper to serve

@zephyr_z9 well that's the problem

@zephyr_z9 My strategy analysis !

as follows📈📈 👇 👇 👇

@scaling01 hence sonnet 5 supposedly soon

@scaling01 I've already got direct access to a stupid person; I don't want to pay for access to stupid models.

@babyzitong @zephyr_z9 very good

@scaling01 cost matters. but to save cost, it's better to use open weights instead of a closed source lab's second-class citizen

@scaling01 it’s primarily due to lack of models that can reliably solve a problem for a domain. no hybrid systems that work well outside of agentic scope where delegation is big.

otherwise all harnesses use small models somewhere for reteieval/compaction help.

@scaling01 Intelligence also allows for them to ultimately make the best cheapest models if they wanted to, through both distillation and recursive self improvement

@scaling01 Composer 2.5 proves you wrong.

@scaling01 intelligence per dollar. It's always about best bang for buck.

@scaling01 I think Sonnet is loved a lot

@scaling01 Buddy's never used workflows, or done analysis over 1000s of documents. A lot of AI applications that are just 'automate some closed task xyz' can fare with smaller models just fine.

@scaling01 I use sonnet a lot. Claude uses sonnet a lot

@scaling01 I love sonnet