Mannas models

Mannas, which launched as NeuroDX, is a Bengaluru lab that releases open-weight EEG foundation models under the Pratibha (formerly MANAS) name and builds the Sphota language model.

Bengaluru, Indiamannas.ai

Releases
4
Open weights
3
Latest
Nov 16, 2026

November 2026

1 release
  1. Small modelProprietary

    Sphota

    Sphota is a 240M-parameter language model from Mannas that keeps an explicit, evolving representation of meaning alongside the token sequence rather than relying on standard autoregressive design alone.

    Modalities
    text → text
    Context
    48K context
    Parameters
    240M
    • Trained on roughly 2 trillion tokens, according to Mannas
    • Mannas reported an average of 56.13 across 15 benchmarks, ahead of LFM2-700M at 51.77 in its own evaluation
    • Native context window of 48K tokens
    • Offered through an early-access request form at launch, with no public weights
    Source: Sphota: A Meaning-First Language Model | Mannas

September 2026

2 releases
  1. EmbeddingOpen weights

    MANAS-2 (Pratibha-2)

    MANAS-2, later renamed Pratibha-2, is the flagship EEG foundation model from Mannas, an encoder that combines waveform and spectral-band masked reconstruction with a constraint on reconstructed signal energy.

    Modalities
    time series → embedding
    Parameters
    92.8M
    License
    Apache-2.0
    At launch
    Open weights
    • 92.8M parameters across 22 transformer layers with 512-dimensional embeddings
    • Pretrained on unlabeled clinical EEG from internal sources, TUH EEG and I-CARE
    • Described in the paper MANAS-2: Constrained Reconstruction for EEG Foundation Models, posted to arXiv on September 12, 2026
    • Mannas said it used the model to power its EpiDx and MindDx diagnostic products
    Source: MannasAI/pratibha-2 · Hugging Face
  2. EmbeddingOpen weights

    AXON

    AXON is an EEG foundation model from Mannas that splits attention into a temporal path along each electrode and a spatial path across electrodes, mixed by a learned per-token gate.

    Modalities
    time series → embedding
    Parameters
    118.6M
    License
    Apache-2.0
    At launch
    Open weights
    • Pretrained with masked autoencoding; the released encoder has 118.6M parameters
    • Mannas reported higher mean balanced accuracy than a dense-attention baseline on six EEG tasks
    • Ships through the braindecode library
    • Described in the paper Adaptive Anisotropic Attention for Axis-Structured Signals, posted to arXiv on September 8, 2026
    Source: MannasAI/pratibha-axon · Hugging Face

February 2026

1 release
  1. EmbeddingOpen weights

    MANAS-1

    MANAS-1 is the first EEG foundation model from NeuroDX, now Mannas, a transformer encoder that turns brain recordings into features for diagnosis and brain-computer interface research.

    Modalities
    time series → embedding
    Parameters
    40M
    License
    GPL-3.0
    At launch
    Open weights
    • The launch repository held a small encoder of about 40M parameters with 12 transformer layers and 512-dimensional embeddings
    • Takes EEG at 200 Hz in one-second patches and accepts varying channel counts and recording lengths
    • A medium encoder of about 400M parameters followed on Hugging Face on September 17, 2026
    • First published as NeuroDX/manas-1 and later moved to the MannasAI organization under the Pratibha name
    Source: MannasAI/pratibha-1-small-40M · Hugging Face

Cite and reuse this data

These release records are free to reuse under CC BY 4.0, including commercially. Credit The AI Engineering and link to this page's URL.

Dates are the labs' official announcement dates, and every record links to its source. List fields in the CSV are separated by " | ".

Citation

The AI Engineering Editors. "Mannas models: release history and timeline." The AI Engineering, updated Nov 16, 2026. https://theaiengineering.com/releases/mannas

HTML link

<a href="https://theaiengineering.com/releases/mannas">Mannas models: release history and timeline</a> (The AI Engineering)
Full dataset
1,250 releases from 73 labs
Coverage
to
Updated
License
CC BY 4.0
Publisher
The AI Engineering

Other labs