Breaking

MaruthLabs announces Madhuram-v1, a 151M-parameter model trained on 1T tokens for edge and task-specific AI

October 6, 2026 at 10:18 EDT

Foundation Models
Software Dev & Coding

MaruthLabs has announced Madhuram-v1, a 151 million-parameter language model trained on 1 trillion tokens and positioned for CPU-friendly deployment on phones, laptops and other constrained devices. The company says the new version leads 9 of 17 pretraining benchmarks against LFM2.5-230M, Gemma 3 270M and SmolLM2-135M, while keeping the same compact parameter count as the earlier Madhuram-v0.6.

Oct. 2026 · MaruthLabs

Madhuram-v1 scales training, not size

The 151M-parameter model moves from 600B to 1T training tokens and is aimed at CPU-friendly, task-specific AI for phones, laptops and private deployments.

151M

parameters, unchanged from v0.6

pretraining tokens for v1

9/17

benchmarks led in the comparison set

119

v0.6 decode tok/s on phone CPU

Training budget increased

Same 151M-parameter footprint, with more pretraining tokens.

600Bv0.6

1Tv1

The training-token increase is the headline change: v1 keeps the compact size while moving to a larger pretraining run.

What improves

Reported gains include MMLU, CommonSenseQA, MMLU Pro and TriviaQA, plus leading scores on MBPP and HumanEval in the four-model comparison.

What remains limited

Not every benchmark moved upward, and small models still require careful task scoping and fact-checking because they can hallucinate.

Target workflow

Classify→Extract→Route→Deploy locally

The model is framed for narrow, repeatable language tasks where low latency, privacy and low inference cost matter.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.