Udgam Labs released Laghu on October 22, an English research chat model with about 152 million parameters. Its authors report $146.26 for the main pretraining run and approximately $194 for the whole project. Those are reported computing costs for a particular experiment, rather than the complete cost of creating and maintaining an arbitrary AI product.
The recommended quantised file is approximately 163 MB, and the chat configuration supports up to 16K tokens of context. Its small size is interesting for local experiments, but is insufficient reason to treat it as a dependable general-purpose assistant. Its ability to perform the chosen task needs to be established first.
Results include the weaknesses
The published laghu-evals dataset provides results and evaluation limitations. The authors report 7.9% on GSM8K with thinking enabled and say Laghu wins no preregistered category against all three rivals. This is a research release, rather than a demonstrated leader in its class.
Conversation-quality assessment uses another model as the judge. The dataset distinguishes cases with individual outputs from cases where licences permit only aggregate metrics. An external reviewer must therefore establish which numbers can actually be recomputed from the available files.
That transparency is useful because it separates two different stories: what the model can do and what can be checked about it. Publishing a failure does not automatically improve answer quality. It does, however, allow the next team to avoid repeating the same trust-based assessment without access to the evidence.
Beyond the model file
For local deployment, we recommend recording the precise revision, format, context and generation settings. Two applications can load the same model and use it differently. A comparison that does not fix these conditions has difficulty explaining where a difference in results comes from.
A useful first test would include short questions with verifiable answers, follow-up clarifications and cases where the model should acknowledge missing information. Response time and resource consumption should be measured separately. Bulgarian use would require a separate language test; the stated English purpose does not establish Bulgarian quality.
These are pilot suggestions, rather than results of SIExpo testing. We have not trained or run Laghu and do not independently verify the reported costs. The authors' comparisons remain their measurements under their stated conditions.
Openness has several layers
The release includes code, model files, a recipe and surviving checkpoints. The authors' code and weights use Apache 2.0, while data and external components carry separate notices. One general label does not override the terms for each dataset or dependency.
Our editorial assessment is that Laghu's value lies in being an accessible object of study: small enough for practical examination, with visible boundaries around its results. Selecting it for a real service should follow a team's own criteria for accuracy, language, resources and maintenance. A published recipe begins that examination; it does not establish a successful outcome.
*Cover: a conceptual AI illustration.*
*Sources checked: 2026-10-23T16:10:40.049366+00:00.*
