A small team called SupraLabs this week trained a 25M-parameter language model (about one-sixtieth the size of GPT-2) on consumer-grade GPUs (RTX 5060 Ti 16GB + RTX 5060 8GB), claiming results on par with their own 50M predecessor — our judgment: this reads more as a community showcase than a quotable industry signal, but the counter-trend it points to deserves a note.
What This Is
Supra2-Medium-Base is a base model — that is, a foundation that has completed only pretraining, with no instruction tuning (i.e., the additional training pass on QA/dialogue data). It can continue text but cannot be used directly for chat. In the release post, the team included a sample response from the model on "artificial intelligence," and the text is full of disjointed, repetitive sentences. On the question of whether 25M can match 50M, that sample is itself the most direct rebuttal. The team also acknowledged that an instruction-tuned version "may be released in the future"; the current shape is closer to a research artifact.
Industry View
Community reaction is polarized.
Skeptics raise three hard objections: First, comparing against their own previous version does not constitute a meaningful benchmark — benchmarks only have comparative value when they cross teams and scales; second, the public samples have broken grammar and questionable output quality; third, the task complexity a 25M-scale model can handle is inherently limited, sitting at an order-of-magnitude gap from real business scenarios.
Supporters see this as a signal: while OpenAI and Google continue piling on parameters, the open-source community is moving in the opposite direction, exploring a "small and specialized" route. The team has previewed a four-model series ranging from 25M to 100M.
We lean toward the former: today's 25M models have a capability ceiling still far from real enterprise workloads. But tracking "small models running locally" as a long-term direction is reasonable — if that path works out, its significance would rival the parameter-stacking route.
Impact on Regular People
For enterprise IT: In the short term, do not treat these small models as substitutes for large models. But if the cost of running small models on-device continues to fall over the next few years, data-privacy-sensitive scenarios (finance, healthcare, government) may gain another option for local deployment.
For individual professionals: Irrelevant to daily office work at this stage. A usable local AI assistant still depends on the 7B+ model scale; ordinary users need not change any workflow for a 25M model.
For the consumer market: If this direction works out, the biggest winners are end users — future phones and PCs could run useful AI without depending on the cloud, delivering structural improvements in privacy and response speed.