Tasks and inputs
LingSante-Pulse-1.0 is listed for Embeddings, Feature extraction, Time Series Forecasting. It accepts text and produces embedding. Use the publisher’s input format and preprocessing for your chosen deployment.
SyncDev model directory · Profile updated Nov 21, 2026
LingSante-Pulse-1.0 is inclusionAI's MIT-licensed physiological signal encoder that turns ECG, EEG, PPG and other biosignals into 256-dim features.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Pricing conditions: Open weights under MIT; no hosted API.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
Biosignal teams often rebuild a feature extractor for every new dataset: one for ECG, another for sleep EEG, a third for oximetry. LingSante-Pulse-1.0, published by inclusionAI on 20 November 2026, is meant to be the shared front end instead. It is a compact encoder, about 16.7M parameters, that converts windows of physiological waveforms into fixed-size representations you then feed to your own classifier.
It does not diagnose anything. There are no prediction heads, and inclusionAI says so directly.
Signals arrive as [T], [C, T] or [B, C, T] arrays, or as channel dictionaries. Two pieces of metadata are mandatory: the real sampling rate, and a physical identity per channel such as {"modality": "ECG", "lead": "II"}. Everything is resampled to a fixed 128 Hz internally, and channels must be time-aligned.
The release defines 57 channel identities. They span:
Anything outside the list, an IMU stream for example, raises an error rather than being guessed. Numeric channels are normalised with fixed physical constants (SpO₂ centred at 95 with scale 5; heart rate at 70 with scale 30) and clipped. NaN or infinite values count as missing, while zero is a real sample.
| Output | Shape | Meaning |
|---|---|---|
cls |
[B, C×256] |
One 256-dim CLS vector per channel, concatenated |
patch_mean |
[B, C×256] |
Mean over valid patches per channel |
patch_tokens |
[B, P, C×256] |
Optional, every 0.25 s |
patch_valid |
[B, P, C] |
Optional validity mask |
Channels are encoded independently. A twelve-lead ECG therefore yields a 3,072-dimensional vector, which is worth knowing before you size a downstream model.
Official install notes call for a separate inference-code package. No such package is published on PyPI at the time of writing. The code lives in inclusionAI's LingSante-Pulse repository: install it from source with Python 3.11 or 3.12, then call LingSantePulse.from_pretrained(), which pulls the weights from the Hub. If a package does appear later, prefer it; until then, the repository is the route that works.
That same repository holds a second, unrelated branch for continuous glucose monitoring, producing 146-dimensional daily representations with up to seven days of history. Its weights ship inside the code, not in this model repo, so treat it as a separate tool.
Freeze the encoder and train a logistic regression on top. That is a fair, cheap test. Compare cls against patch_mean features, and against hand-crafted baselines (heart-rate variability, band powers) you already trust. Check channel mismatch too: does an ECG recorded at 500 Hz behave like the 128 Hz cases? Medical use also brings regulatory questions that a general MIT licence does not answer.
inclusionAI is better known for language models such as Ling 3.2 Flash; its full catalogue is on the inclusionAI provider page. Comparable encoders sit under feature extraction.
Compare the model on the work you need to complete.
LingSante-Pulse-1.0 is listed for Embeddings, Feature extraction, Time Series Forecasting. It accepts text and produces embedding. Use the publisher’s input format and preprocessing for your chosen deployment.
Evaluate retrieval over a labelled query set. Keep the encoder, vector dimensions and index settings consistent.
The weights are downloadable under MIT. Check runtime support and size the complete model and working context for your expected concurrency.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
Qwen
Qwen/Qwen3-Embedding-8B is a source-linked Hugging Face repository assigned to Feature extraction. Its recorded task interface accepts text and produces embeddings.
NVIDIA Research
Hugging Face repository for Feature extraction.
Qwen
Qwen/Qwen3-Embedding-4B is a source-linked Hugging Face repository assigned to Feature extraction. Its recorded task interface accepts text and produces embeddings.
Qwen
Qwen/Qwen3-Embedding-0.6B is a source-linked Hugging Face repository assigned to Feature extraction. Its recorded task interface accepts text and produces embeddings.
BAAI
BAAI/bge-multilingual-gemma2 is a source-linked Hugging Face repository assigned to Feature extraction. Its recorded task interface accepts text and produces embeddings.
Ai2 (Allen Institute for AI)
Hugging Face repository for Feature extraction.